class UnmergedEdgeExampleBuilder(ExampleBuilder):
def __init__(self, style=["typed","directed","headsOnly"], length=None, types=[], featureSet=None, classSet=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert( classSet.getId("neg") == 1 )
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
self.styles = style
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
self.pathLengths = length
assert(self.pathLengths == None)
self.types = types
if "random" in self.styles:
from FeatureBuilders.RandomFeatureBuilder import RandomFeatureBuilder
self.randomFeatureBuilder = RandomFeatureBuilder(self.featureSet)
#self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls, input, output, parse, tokenization, style, idFileTag=None):
classSet, featureSet = cls.getIdSets(idFileTag)
if style == None:
e = UnmergedEdgeExampleBuilder(classSet=classSet, featureSet=featureSet)
else:
e = UnmergedEdgeExampleBuilder(style=style, classSet=classSet, featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
print e.classSet.Ids
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def filterEdgesByType(self, edges, typesToInclude):
if len(typesToInclude) == 0:
return edges
edgesToKeep = []
for edge in edges:
if edge.get("type") in typesToInclude:
edgesToKeep.append(edge)
return edgesToKeep
def getCategoryName(self, sentenceGraph, e1, e2, directed=True):
# Dummies are potential entities that do not exist in the
# training data. If both entities of an interaction are dummies
# it can't exist in the training data and is therefore a negative
if e1[2] or e2[2]:
return "neg"
e1 = e1[0]
e2 = e2[0]
interactions = sentenceGraph.getInteractions(e1, e2)
if not directed:
interactions.extend(sentenceGraph.getInteractions(e2, e1))
types = set()
for interaction in interactions:
types.add(interaction.attrib["type"])
types = list(types)
types.sort()
categoryName = ""
for name in types:
if categoryName != "":
categoryName += "---"
categoryName += name
if categoryName != "":
return categoryName
else:
return "neg"
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def isPotentialGeniaInteraction(self, e1, e2):
if e1.get("isName") == "True":
return False
else:
return True
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def getInteractionEdgeLengths(self, sentenceGraph, paths):
"""
Return dependency and linear length of all interaction edges
(measured between the two tokens).
"""
interactionLengths = {}
for interaction in sentenceGraph.interactions:
# Calculated interaction edge dep and lin length
e1 = sentenceGraph.entitiesById[interaction.get("e1")]
e2 = sentenceGraph.entitiesById[interaction.get("e2")]
t1 = sentenceGraph.entityHeadTokenByEntity[e1]
t2 = sentenceGraph.entityHeadTokenByEntity[e2]
# Get dep path length
if t1 != t2 and paths.has_key(t1) and paths[t1].has_key(t2):
pathLength = len(paths[t1][t2])
else: # no dependencyPath
pathLength = 999999 # more than any real path
# Linear distance
t1Pos = -1
t2Pos = -1
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == t1:
t1Pos = i
if t2Pos != -1:
break
if sentenceGraph.tokens[i] == t2:
t2Pos = i
if t1Pos != -1:
break
linLength = abs(t1Pos - t2Pos)
interactionLengths[interaction] = (pathLength, linLength)
return interactionLengths
def getPrecedenceLevels(self, sentenceGraph, paths):
"""
Get overlapping entity precedence
"""
interactionLengths = self.getInteractionEdgeLengths(sentenceGraph, paths)
interactionsByEntity = {} # Convenience mapping
entityPrecedenceValues = {}
for entity in sentenceGraph.entities:
interactionsByEntity[entity] = []
eId = entity.get("id")
# Add access to interactions
argDepDist = 0 # Sum of lengths of shortest paths
argLinDist = 0 # Sum of linear distances
for interaction in sentenceGraph.interactions:
if interaction.get("e1") == eId: # An argument of the entity defined by the node
interactionsByEntity[entity].append(interaction)
argDepDist += interactionLengths[interaction][0]
argLinDist += interactionLengths[interaction][1]
# Store precedence counts (num args, sum of dep lengths, sum of lin lengths)
entityPrecedenceValues[entity] = (len(interactionsByEntity), argDepDist, argLinDist, entity)
# Determine level of entity from precedence counts
levelByEntity = {} # slot number
#levelByInteraction = {} # slot number of parent node
# There is one slot group per token, per type
for token in sentenceGraph.tokens: # per token
entitiesByType = {}
for entity in sentenceGraph.tokenIsEntityHead[token]: # per type
if entity.get("isName") == "True": # Names can never have duplicates
assert not levelByEntity.has_key(entity)
levelByEntity[entity] = 0
continue
eType = entity.get("type")
if eType == "neg":
continue
if not entitiesByType.has_key(eType):
entitiesByType[eType] = []
entitiesByType[eType].append(entity)
for eType in sorted(entitiesByType.keys()):
# Slot ordering by precedence
sortedEntities = []
for entity in entitiesByType[eType]:
sortedEntities.append(entityPrecedenceValues[entity])
sortedEntities.sort(compareEntityPrecedence)
level = 0
for precedenceTuple in sortedEntities:
entity = precedenceTuple[3]
assert not levelByEntity.has_key(entity)
levelByEntity[entity] = level
# Interactions have the same slot as their parent entity
#for interaction in interactionsByEntity[entity]:
# assert not levelByInteraction.has_key(interaction)
# levelByInteraction[interaction] = level
level += 1
return levelByEntity#, levelByInteraction
def buildExamples(self, sentenceGraph):
examples = []
exampleIndex = 0
#undirected = sentenceGraph.getUndirectedDependencyGraph()
undirected = self.nxMultiDiGraphToUndirected(sentenceGraph.dependencyGraph)
##undirected = sentenceGraph.dependencyGraph.to_undirected()
###undirected = NX10.MultiGraph(sentenceGraph.dependencyGraph) This didn't work
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
# Determine overlapping entity precedence
#levelByEntity, levelByInteraction = self.getPrecedenceLevels(sentenceGraph, paths)
levelByEntity = self.getPrecedenceLevels(sentenceGraph, paths)
entities = []
# There is one entity group for each token, for each type of entity
for token in sentenceGraph.tokens: # per token
entitiesByType = {}
for entity in sentenceGraph.tokenIsEntityHead[token]: # per type
if entity.get("isName") == "True": # Names can never have duplicates
entities.append( (entity, 0, False) )
continue
eType = entity.get("type")
if eType == "neg":
continue
if not entitiesByType.has_key(eType):
entitiesByType[eType] = []
entitiesByType[eType].append(entity)
# Create slot groups for tokens for which exists at least one entity
eTypes = sorted(entitiesByType.keys())
if len(eTypes) == 0:
continue
# Create slot groups and insert GS data there
for eType in eTypes:
# Use first entity of a type as the dummy entity for unfilled slots
dummyEntity = entitiesByType[eType][0]
# Define entity slots
entityGroup = [None, None, None, None]
#entityGroup = [None, None]
# Insert existing entities into slots
for entity in entitiesByType[eType]:
if levelByEntity.has_key(entity):
level = levelByEntity[entity]
if level < len(entityGroup):
entityGroup[level] = (entity, level, False)
# Create dummies for potential entities
for i in range(len(entityGroup)):
if entityGroup[i] == None:
entityGroup[i] = (dummyEntity, i, True)
# Put all slots into one potential entity list
#print entityGroup
for e in entityGroup:
entities.append(e)
# Generate examples based on interactions between entities
for i in range(len(entities)-1):
for j in range(i+1,len(entities)):
eI = entities[i][0]
eJ = entities[j][0]
tI = sentenceGraph.entityHeadTokenByEntity[eI]
tJ = sentenceGraph.entityHeadTokenByEntity[eJ]
# define forward example
categoryName = self.getCategoryName(sentenceGraph, entities[i], entities[j], True)
if (not "genia_limits" in self.styles) or self.isPotentialGeniaInteraction(eI, eJ):
examples.append( self.buildExample(tI, tJ, paths, sentenceGraph, categoryName, exampleIndex, entities[i], entities[j]) )
exampleIndex += 1
# define reverse
categoryName = self.getCategoryName(sentenceGraph, entities[j], entities[i], True)
if (not "genia_limits" in self.styles) or self.isPotentialGeniaInteraction(eJ, eI):
examples.append( self.buildExample(tJ, tI, paths, sentenceGraph, categoryName, exampleIndex, entities[j], entities[i]) )
exampleIndex += 1
return examples
def buildExample(self, token1, token2, paths, sentenceGraph, categoryName, exampleIndex, e1=None, e2=None):
entity1=e1[0]
entity2=e2[0]
# define features
features = {}
features[self.featureSet.getId("gov_level")] = e1[1]
features[self.featureSet.getId("gov_level_"+str(e1[1]))] = 1
features[self.featureSet.getId("dep_level")] = e2[1]
features[self.featureSet.getId("dep_level_"+str(e2[1]))] = 1
features[self.featureSet.getId("level_pair_"+str(e1[1])+"_"+str(e2[1]))] = 1
if True: #token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
if token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
path = paths[token1][token2]
else:
path = [token1, token2]
assert(self.pathLengths == None)
if self.pathLengths == None or len(path)-1 in self.pathLengths:
if not "no_dependency" in self.styles:
if token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
edges = self.multiEdgeFeatureBuilder.getEdges(sentenceGraph.dependencyGraph, path)
else:
edges = None
if "entity_type" in self.styles:
features[self.featureSet.getId("e1_"+entity1.attrib["type"])] = 1
features[self.featureSet.getId("e2_"+entity2.attrib["type"])] = 1
features[self.featureSet.getId("distance_"+str(len(path)))] = 1
if not "no_dependency" in self.styles:
self.multiEdgeFeatureBuilder.setFeatureVector(features, entity1, entity2)
#self.multiEdgeFeatureBuilder.buildStructureFeatures(sentenceGraph, paths) # remove for fast
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
if not "disable_terminus_features" in self.styles:
self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(4, path, edges, sentenceGraph) # remove for fast
#self.buildEdgeCombinations(path, edges, sentenceGraph, features) # remove for fast
#if edges != None:
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[0], edges[0][1]+edges[1][0], "t1", sentenceGraph) # remove for fast
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[-1], edges[len(path)-1][len(path)-2]+edges[len(path)-2][len(path)-1], "t2", sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(path, edges, sentenceGraph)
self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None)
if not "no_linear" in self.styles:
self.tokenFeatureBuilder.setFeatureVector(features)
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == token1:
token1Index = i
if sentenceGraph.tokens[i] == token2:
token2Index = i
linearPreTag = "linfw_"
if token1Index > token2Index:
token1Index, token2Index = token2Index, token1Index
linearPreTag = "linrv_"
self.tokenFeatureBuilder.buildLinearOrderFeatures(token1Index, sentenceGraph, 2, 2, preTag="linTok1")
self.tokenFeatureBuilder.buildLinearOrderFeatures(token2Index, sentenceGraph, 2, 2, preTag="linTok2")
# Before, middle, after
# self.tokenFeatureBuilder.buildTokenGrams(0, token1Index-1, sentenceGraph, "bf")
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, "bw")
# self.tokenFeatureBuilder.buildTokenGrams(token2Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, "af")
# before-middle, middle, middle-after
# self.tokenFeatureBuilder.buildTokenGrams(0, token2Index-1, sentenceGraph, linearPreTag+"bf", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, linearPreTag+"bw", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, linearPreTag+"af", max=2)
self.tokenFeatureBuilder.setFeatureVector(None)
if "random" in self.styles:
self.randomFeatureBuilder.setFeatureVector(features)
self.randomFeatureBuilder.buildRandomFeatures(100, 0.01)
self.randomFeatureBuilder.setFeatureVector(None)
if "genia_limits" in self.styles:
e1Type = entity1.get("type")
e2Type = entity2.get("type")
assert(entity1.get("isName") == "False")
if entity2.get("isName") == "True":
features[self.featureSet.getId("GENIA_target_protein")] = 1
else:
features[self.featureSet.getId("GENIA_nested_event")] = 1
if e1Type.find("egulation") != -1: # leave r out to avoid problems with capitalization
if entity2.get("isName") == "True":
features[self.featureSet.getId("GENIA_regulation_of_protein")] = 1
else:
features[self.featureSet.getId("GENIA_regulation_of_event")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
path = [token1, token2]
# define extra attributes
if int(path[0].attrib["id"].split("_")[-1]) < int(path[-1].attrib["id"].split("_")[-1]):
#extra = {"xtype":"edge","type":"i","t1":path[0],"t2":path[-1]}
extra = {"xtype":"ue","type":"i","t1":path[0].get("id"),"t2":path[-1].get("id")}
extra["deprev"] = False
else:
#extra = {"xtype":"edge","type":"i","t1":path[-1],"t2":path[0]}
extra = {"xtype":"ue","type":"i","t1":path[-1].get("id"),"t2":path[0].get("id")}
extra["deprev"] = True
if entity1 != None:
extra["e1"] = entity1.get("id")
extra["l1"] = str(e1[1])
extra["d1"] = str(e1[2])[0] # is a dummy node (an entity not in existing triggers)
if entity2 != None:
extra["e2"] = entity2.get("id")
extra["l2"] = str(e2[1])
extra["d2"] = str(e2[2])[0] # is a dummy node (an entity not in existing triggers)
extra["categoryName"] = categoryName
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
if "binary" in self.styles:
if categoryName != "neg":
category = 1
else:
category = -1
categoryName = "i"
else:
category = self.classSet.getId(categoryName)
return (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra)
class EventExampleBuilder(ExampleBuilder):
def __init__(self,
style=["typed", "directed", "headsOnly"],
length=None,
types=[],
featureSet=None,
classSet=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert (classSet.getId("neg") == 1)
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
self.styles = style
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
#self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
#if "ontology" in self.styles:
# self.multiEdgeFeatureBuilder.ontologyFeatureBuilder = BioInferOntologyFeatureBuilder(self.featureSet)
self.pathLengths = length
assert (self.pathLengths == None)
self.types = types
#self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls, input, output, parse, tokenization, style, idFileTag=None):
classSet, featureSet = cls.getIdSets(idFileTag)
e = EventExampleBuilder(style=style,
classSet=classSet,
featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(
sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def isPotentialGeniaInteraction(self, e1, e2):
if e1.get("isName") == "True" and e2.get("isName") == "True":
return False
elif e1.get("isName") == "True" and e2.get("isName") == "False":
return False
else:
return True
def getArgumentEntities(self, sentenceGraph, entityNode):
eId = entityNode.get("id")
assert (eId != None)
themeNodes = []
causeNodes = []
for edge in sentenceGraph.interactions:
if edge.get("e1") == eId:
edgeType = edge.get("type")
assert (edgeType in ["Theme", "Cause"]), edgeType
if edgeType == "Theme":
themeNodes.append(
sentenceGraph.entitiesById[edge.get("e2")])
elif edgeType == "Cause":
causeNodes.append(
sentenceGraph.entitiesById[edge.get("e2")])
return themeNodes, causeNodes
def makeGSEvents(self, sentenceGraph):
self.gsEvents = {} # [token]->[event-type]->[1-n argument sets]
for token in sentenceGraph.tokens:
self.gsEvents[token] = {}
for entity in sentenceGraph.entities:
if entity.get("type") == "neg":
continue
eId = entity.get("id")
eType = entity.get("type")
arguments = set()
for interaction in sentenceGraph.interactions:
if interaction.get("e1") == eId:
arguments.add(
(interaction.get("type"), interaction.get("e2")))
eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
if not self.gsEvents[eHeadToken].has_key(eType):
self.gsEvents[eHeadToken][eType] = []
self.gsEvents[eHeadToken][eType].append(arguments)
def isGSEvent(self, sentenceGraph, entity, themeNodes, causeNodes):
eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
eType = entity.get("type")
if not self.gsEvents[eHeadToken].has_key(eType):
return False
argumentSet = set()
for themeNode in themeNodes:
if themeNode != None:
argumentSet.add(("Theme", themeNode.get("id")))
for causeNode in causeNodes:
if causeNode != None:
argumentSet.add(("Cause", causeNode.get("id")))
if argumentSet in self.gsEvents[eHeadToken][eType]:
return True
else:
return False
# def isEvent(self, sentenceGraph, eventNode, themeNodes, causeNodes):
# goldThemeNodes, goldCauseNodes = self.getArgumentEntities(sentenceGraph, eventNode)
# for node in themeNodes:
# if node != None and node not in goldThemeNodes:
# return False
# for node in causeNodes:
# if node != None and node not in goldCauseNodes:
# return False
# return True
def buildExamples(self, sentenceGraph):
self.makeGSEvents(sentenceGraph)
eventNodes = []
nameNodes = []
for entity in sentenceGraph.entities:
if entity.get("type") == "neg":
continue
if entity.get("isName") == "True":
nameNodes.append(entity)
else:
eventNodes.append(entity)
allNodes = eventNodes + nameNodes
examples = []
exampleIndex = 0
undirected = sentenceGraph.dependencyGraph.to_undirected()
paths = NX.all_pairs_shortest_path(undirected, cutoff=999)
for eventNode in eventNodes:
eventType = eventNode.get("type")
if eventType in [
"Gene_expression", "Transcription", "Protein_catabolism",
"Localization", "Phosphorylation"
]:
for nameNode in nameNodes:
if self.isPotentialGeniaInteraction(eventNode, nameNode):
examples.append(
self.buildExample(exampleIndex, sentenceGraph,
paths, eventNode, nameNode))
exampleIndex += 1
elif eventType in [
"Regulation", "Positive_regulation", "Negative_regulation"
]:
combinations = combine.combine(allNodes + [None],
allNodes + [None])
for combination in combinations:
if combination[0] == combination[1]:
continue
if combination[0] == eventNode or combination[
1] == eventNode:
continue
if combination[
0] != None and not self.isPotentialGeniaInteraction(
eventNode, combination[0]):
continue
if combination[
1] != None and not self.isPotentialGeniaInteraction(
eventNode, combination[1]):
continue
examples.append(
self.buildExample(exampleIndex, sentenceGraph, paths,
eventNode, combination[0],
combination[1]))
exampleIndex += 1
elif eventType in ["Binding"]:
continue
else:
assert False, eventType
self.gsEvents = None
return examples
def buildExample(self,
exampleIndex,
sentenceGraph,
paths,
eventNode,
themeNode,
causeNode=None):
features = {}
if self.isGSEvent(sentenceGraph, eventNode, [themeNode], [causeNode]):
category = self.classSet.getId("pos")
else:
category = self.classSet.getId("neg")
if themeNode != None:
self.buildArgumentFeatures(sentenceGraph, paths, features,
eventNode, themeNode, "theme_")
if causeNode != None:
self.buildArgumentFeatures(sentenceGraph, paths, features,
eventNode, causeNode, "cause_")
# Common features
# eventType = eventNode.get("type")
# e2Type = entity2.get("type")
# assert(entity1.get("isName") == "False")
# if entity2.get("isName") == "True":
# features[self.featureSet.getId("GENIA_target_protein")] = 1
# else:
# features[self.featureSet.getId("GENIA_nested_event")] = 1
# if e1Type.find("egulation") != -1: # leave r out to avoid problems with capitalization
# if entity2.get("isName") == "True":
# features[self.featureSet.getId("GENIA_regulation_of_protein")] = 1
# else:
# features[self.featureSet.getId("GENIA_regulation_of_event")] = 1
# define extra attributes
extra = {"xtype": "trigger-event", "type": eventNode.get("type")}
extra["e"] = eventNode.get("id")
eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
extra["et"] = eventToken.get("id")
if themeNode != None:
extra["t"] = themeNode.get("id")
themeToken = sentenceGraph.entityHeadTokenByEntity[themeNode]
extra["tt"] = themeToken.get("id")
if causeNode != None:
extra["c"] = causeNode.get("id")
causeToken = sentenceGraph.entityHeadTokenByEntity[causeNode]
extra["ct"] = causeToken.get("id")
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
#assert (category == 1 or category == -1)
return (sentenceGraph.getSentenceId() + ".x" + str(exampleIndex),
category, features, extra)
def buildArgumentFeatures(self, sentenceGraph, paths, features, eventNode,
argNode, tag):
eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
argToken = sentenceGraph.entityHeadTokenByEntity[argNode]
if eventToken != argToken and paths.has_key(
eventToken) and paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
edges = self.multiEdgeFeatureBuilder.getEdges(
sentenceGraph.dependencyGraph, path)
else:
path = [eventToken, argToken]
edges = None
self.multiEdgeFeatureBuilder.tag = tag
self.multiEdgeFeatureBuilder.setFeatureVector(features, eventNode,
argNode)
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
if not "disable_terminus_features" in self.styles:
self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(
path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(
path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(
2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
4, path, edges, sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(
path, edges, sentenceGraph)
self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None)
self.multiEdgeFeatureBuilder.tag = ""
class EventExampleBuilder(ExampleBuilder):
def __init__(self, style=["typed", "directed", "headsOnly"], length=None, types=[], featureSet=None, classSet=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert classSet.getId("neg") == 1
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
self.styles = style
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
# self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
# if "ontology" in self.styles:
# self.multiEdgeFeatureBuilder.ontologyFeatureBuilder = BioInferOntologyFeatureBuilder(self.featureSet)
self.pathLengths = length
assert self.pathLengths == None
self.types = types
# self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls, input, output, parse, tokenization, style, idFileTag=None):
classSet, featureSet = cls.getIdSets(idFileTag)
e = EventExampleBuilder(style=style, classSet=classSet, featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >>sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def isPotentialGeniaInteraction(self, e1, e2):
if e1.get("isName") == "True" and e2.get("isName") == "True":
return False
elif e1.get("isName") == "True" and e2.get("isName") == "False":
return False
else:
return True
def getArgumentEntities(self, sentenceGraph, entityNode):
eId = entityNode.get("id")
assert eId != None
themeNodes = []
causeNodes = []
for edge in sentenceGraph.interactions:
if edge.get("e1") == eId:
edgeType = edge.get("type")
assert edgeType in ["Theme", "Cause"], edgeType
if edgeType == "Theme":
themeNodes.append(sentenceGraph.entitiesById[edge.get("e2")])
elif edgeType == "Cause":
causeNodes.append(sentenceGraph.entitiesById[edge.get("e2")])
return themeNodes, causeNodes
def makeGSEvents(self, sentenceGraph):
self.gsEvents = {} # [token]->[event-type]->[1-n argument sets]
for token in sentenceGraph.tokens:
self.gsEvents[token] = {}
for entity in sentenceGraph.entities:
if entity.get("type") == "neg":
continue
eId = entity.get("id")
eType = entity.get("type")
arguments = set()
for interaction in sentenceGraph.interactions:
if interaction.get("e1") == eId:
arguments.add((interaction.get("type"), interaction.get("e2")))
eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
if not self.gsEvents[eHeadToken].has_key(eType):
self.gsEvents[eHeadToken][eType] = []
self.gsEvents[eHeadToken][eType].append(arguments)
def isGSEvent(self, sentenceGraph, entity, themeNodes, causeNodes):
eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
eType = entity.get("type")
if not self.gsEvents[eHeadToken].has_key(eType):
return False
argumentSet = set()
for themeNode in themeNodes:
if themeNode != None:
argumentSet.add(("Theme", themeNode.get("id")))
for causeNode in causeNodes:
if causeNode != None:
argumentSet.add(("Cause", causeNode.get("id")))
if argumentSet in self.gsEvents[eHeadToken][eType]:
return True
else:
return False
# def isEvent(self, sentenceGraph, eventNode, themeNodes, causeNodes):
# goldThemeNodes, goldCauseNodes = self.getArgumentEntities(sentenceGraph, eventNode)
# for node in themeNodes:
# if node != None and node not in goldThemeNodes:
# return False
# for node in causeNodes:
# if node != None and node not in goldCauseNodes:
# return False
# return True
def buildExamples(self, sentenceGraph):
self.makeGSEvents(sentenceGraph)
eventNodes = []
nameNodes = []
for entity in sentenceGraph.entities:
if entity.get("type") == "neg":
continue
if entity.get("isName") == "True":
nameNodes.append(entity)
else:
eventNodes.append(entity)
allNodes = eventNodes + nameNodes
examples = []
exampleIndex = 0
undirected = sentenceGraph.dependencyGraph.to_undirected()
paths = NX.all_pairs_shortest_path(undirected, cutoff=999)
for eventNode in eventNodes:
eventType = eventNode.get("type")
if eventType in [
"Gene_expression",
"Transcription",
"Protein_catabolism",
"Localization",
"Phosphorylation",
]:
for nameNode in nameNodes:
if self.isPotentialGeniaInteraction(eventNode, nameNode):
examples.append(self.buildExample(exampleIndex, sentenceGraph, paths, eventNode, nameNode))
exampleIndex += 1
elif eventType in ["Regulation", "Positive_regulation", "Negative_regulation"]:
combinations = combine.combine(allNodes + [None], allNodes + [None])
for combination in combinations:
if combination[0] == combination[1]:
continue
if combination[0] == eventNode or combination[1] == eventNode:
continue
if combination[0] != None and not self.isPotentialGeniaInteraction(eventNode, combination[0]):
continue
if combination[1] != None and not self.isPotentialGeniaInteraction(eventNode, combination[1]):
continue
examples.append(
self.buildExample(exampleIndex, sentenceGraph, paths, eventNode, combination[0], combination[1])
)
exampleIndex += 1
elif eventType in ["Binding"]:
continue
else:
assert False, eventType
self.gsEvents = None
return examples
def buildExample(self, exampleIndex, sentenceGraph, paths, eventNode, themeNode, causeNode=None):
features = {}
if self.isGSEvent(sentenceGraph, eventNode, [themeNode], [causeNode]):
category = self.classSet.getId("pos")
else:
category = self.classSet.getId("neg")
if themeNode != None:
self.buildArgumentFeatures(sentenceGraph, paths, features, eventNode, themeNode, "theme_")
if causeNode != None:
self.buildArgumentFeatures(sentenceGraph, paths, features, eventNode, causeNode, "cause_")
# Common features
# eventType = eventNode.get("type")
# e2Type = entity2.get("type")
# assert(entity1.get("isName") == "False")
# if entity2.get("isName") == "True":
# features[self.featureSet.getId("GENIA_target_protein")] = 1
# else:
# features[self.featureSet.getId("GENIA_nested_event")] = 1
# if e1Type.find("egulation") != -1: # leave r out to avoid problems with capitalization
# if entity2.get("isName") == "True":
# features[self.featureSet.getId("GENIA_regulation_of_protein")] = 1
# else:
# features[self.featureSet.getId("GENIA_regulation_of_event")] = 1
# define extra attributes
extra = {"xtype": "trigger-event", "type": eventNode.get("type")}
extra["e"] = eventNode.get("id")
eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
extra["et"] = eventToken.get("id")
if themeNode != None:
extra["t"] = themeNode.get("id")
themeToken = sentenceGraph.entityHeadTokenByEntity[themeNode]
extra["tt"] = themeToken.get("id")
if causeNode != None:
extra["c"] = causeNode.get("id")
causeToken = sentenceGraph.entityHeadTokenByEntity[causeNode]
extra["ct"] = causeToken.get("id")
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
# assert (category == 1 or category == -1)
return (sentenceGraph.getSentenceId() + ".x" + str(exampleIndex), category, features, extra)
def buildArgumentFeatures(self, sentenceGraph, paths, features, eventNode, argNode, tag):
eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
argToken = sentenceGraph.entityHeadTokenByEntity[argNode]
if eventToken != argToken and paths.has_key(eventToken) and paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
edges = self.multiEdgeFeatureBuilder.getEdges(sentenceGraph.dependencyGraph, path)
else:
path = [eventToken, argToken]
edges = None
self.multiEdgeFeatureBuilder.tag = tag
self.multiEdgeFeatureBuilder.setFeatureVector(features, eventNode, argNode)
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
if not "disable_terminus_features" in self.styles:
self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(4, path, edges, sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(path, edges, sentenceGraph)
self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None)
self.multiEdgeFeatureBuilder.tag = ""
class Round2TriggerExampleBuilder(ExampleBuilder):
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def getPredictionStrength(self, element):
eType = element.get("type")
predictions = element.get("predictions")
if predictions == None:
return 0
predictions = predictions.split(",")
for prediction in predictions:
predClass, predStrength = prediction.split(":")
if predClass == eType:
predStrength = float(predStrength)
return predStrength
return 0
def getInteractionEdgeLengths(self, sentenceGraph, paths):
"""
Return dependency and linear length of all interaction edges
(measured between the two tokens).
"""
interactionLengths = {}
for interaction in sentenceGraph.interactions:
# Calculated interaction edge dep and lin length
e1 = sentenceGraph.entitiesById[interaction.get("e1")]
e2 = sentenceGraph.entitiesById[interaction.get("e2")]
t1 = sentenceGraph.entityHeadTokenByEntity[e1]
t2 = sentenceGraph.entityHeadTokenByEntity[e2]
# Get dep path length
if t1 != t2 and paths.has_key(t1) and paths[t1].has_key(t2):
pathLength = len(paths[t1][t2])
else: # no dependencyPath
pathLength = 999999 # more than any real path
# Linear distance
t1Pos = -1
t2Pos = -1
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == t1:
t1Pos = i
if t2Pos != -1:
break
if sentenceGraph.tokens[i] == t2:
t2Pos = i
if t1Pos != -1:
break
linLength = abs(t1Pos - t2Pos)
interactionLengths[interaction] = (interaction, pathLength,
linLength, t2Pos)
return interactionLengths
def __init__(self,
style=None,
classSet=None,
featureSet=None,
gazetteerFileName=None,
skiplist=None):
if classSet == None:
classSet = IdSet(1)
assert (classSet.getId("neg") == 1)
if featureSet == None:
featureSet = IdSet()
ExampleBuilder.__init__(self, classSet, featureSet)
#gazetteerFileName="/usr/share/biotext/GeniaChallenge/SharedTaskTriggerTest/gazetteer-train"
if gazetteerFileName != None:
self.gazetteer = Gazetteer.loadGztr(gazetteerFileName)
print >> sys.stderr, "Loaded gazetteer from", gazetteerFileName
else:
print >> sys.stderr, "No gazetteer loaded"
self.gazetteer = None
self.styles = style
self.skiplist = set()
if skiplist != None:
f = open(skiplist, "rt")
for line in f.readlines():
self.skiplist.add(line.strip())
f.close()
self.styles = [
"trigger_features", "typed", "directed", "no_linear", "entities",
"genia_limits", "noMasking", "maxFeatures"
]
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if "graph_kernel" in self.styles:
from FeatureBuilders.GraphKernelFeatureBuilder import GraphKernelFeatureBuilder
self.graphKernelFeatureBuilder = GraphKernelFeatureBuilder(
self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.triggerFeatureBuilder = TriggerFeatureBuilder(self.featureSet)
@classmethod
def run(cls,
input,
gold,
output,
parse,
tokenization,
style,
idFileTag=None,
append=False):
"""
An interface for running the example builder without needing to create a class
"""
classSet, featureSet = cls.getIdSets(idFileTag)
if style != None:
e = Round2TriggerExampleBuilder(style=style,
classSet=classSet,
featureSet=featureSet)
else:
e = Round2TriggerExampleBuilder(classSet=classSet,
featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
if gold != None:
goldSentences = cls.getSentences(gold, parse, tokenization)
else:
goldSentences = None
e.buildExamplesForSentences(sentences,
goldSentences,
output,
idFileTag,
append=append)
def buildExamplesForSentences(self,
sentences,
goldSentences,
output,
idFileTag=None,
append=False):
examples = []
counter = ProgressCounter(len(sentences), "Build examples")
if append:
outfile = open(output, "at")
else:
outfile = open(output, "wt")
exampleCount = 0
for i in range(len(sentences)):
sentence = sentences[i]
goldSentence = [None]
if goldSentences != None:
goldSentence = goldSentences[i]
counter.update(
1, "Building examples (" + sentence[0].getSentenceId() + "): ")
examples = self.buildExamples(sentence[0],
goldSentence[0],
append=append)
exampleCount += len(examples)
examples = self.preProcessExamples(examples)
ExampleUtils.appendExamples(examples, outfile)
outfile.close()
print >> sys.stderr, "Examples built:", exampleCount
print >> sys.stderr, "Features:", len(self.featureSet.getNames())
#IF LOCAL
if self.exampleStats.getExampleCount() > 0:
self.exampleStats.printStats()
#ENDIF
# Save Ids
if idFileTag != None:
print >> sys.stderr, "Saving class names to", idFileTag + ".class_names"
self.classSet.write(idFileTag + ".class_names")
print >> sys.stderr, "Saving feature names to", idFileTag + ".feature_names"
self.featureSet.write(idFileTag + ".feature_names")
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def getMergedEntityType(self, entities):
"""
If a single token belongs to multiple entities of different types,
a new, composite type is defined. This type is the alphabetically
ordered types of these entities joined with '---'.
"""
types = set()
for entity in entities:
types.add(entity.get("type"))
types = list(types)
types.sort()
typeString = ""
for type in types:
if type == "Protein" and "all_tokens" in self.styles:
continue
if typeString != "":
typeString += "---"
typeString += type
if typeString == "":
return "neg"
if "limit_merged_types" in self.styles:
if typeString.find("---") != -1:
if typeString == "Gene_expression---Positive_regulation":
return typeString
else:
return typeString.split("---")[0]
else:
return typeString
return typeString
def getTokenFeatures(self, token, sentenceGraph):
"""
Returns a list of features based on the attributes of a token.
These can be used to define more complex features.
"""
# These features are cached when this method is first called
# for a token.
if self.tokenFeatures.has_key(token):
return self.tokenFeatures[token]
tokTxt = sentenceGraph.getTokenText(token)
features = {}
features["_txt_" + tokTxt] = 1
# F 69.35 -> 68.22
#normalizedText = tokTxt.replace("-","").replace("/","").replace(",","").replace("\\","").replace(" ","").lower()
#features["_norTxt_"+normalizedText]=1
#features["_norStem_" + PorterStemmer.stem(normalizedText)]=1
features["_POS_" + token.get("POS")] = 1
if sentenceGraph.tokenIsName[token]:
features["_isName"] = 1
for entity in sentenceGraph.tokenIsEntityHead[token]:
if entity.get("isName") == "True":
features["_annType_" + entity.get("type")] = 1
# Filip's gazetteer based features (can be used separately from exclude_gazetteer)
if "gazetteer_features" in self.styles:
tokTxtLower = tokTxt.lower()
if "stem_gazetteer" in self.styles:
tokTxtLower = PorterStemmer.stem(tokTxtLower)
if self.gazetteer and tokTxtLower in self.gazetteer:
for label, weight in self.gazetteer[tokTxtLower].items():
features["_knownLabel_" +
label] = weight # 1 performs slightly worse
self.tokenFeatures[token] = features
return features
def buildLinearOrderFeatures(self, sentenceGraph, index, tag, features):
"""
Linear features are built by marking token features with a tag
that defines their relative position in the linear order.
"""
tag = "linear_" + tag
for tokenFeature, w in self.getTokenFeatures(
sentenceGraph.tokens[index], sentenceGraph).iteritems():
features[self.featureSet.getId(tag + tokenFeature)] = w
def buildExamples(self, sentenceGraph, goldGraph, append=False):
examples = self.buildExamplesInner(sentenceGraph, goldGraph)
entityCounts = {}
exampleCounts = {}
for entity in sentenceGraph.entities:
eType = entity.get("type")
if eType == "Protein":
continue
if not entityCounts.has_key(eType):
entityCounts[eType] = 0
exampleCounts[eType] = 0
entityCounts[eType] += 1
for example in examples:
eTypes = self.classSet.getName(example[1]).split("---")
for eType in eTypes:
if not exampleCounts.has_key(eType):
exampleCounts[eType] = 0
exampleCounts[eType] += 1
#for key in sorted(entityCounts.keys()):
# if entityCounts[key] != exampleCounts[key]:
# print >> sys.stderr, "Warning, sentence", sentenceGraph.getSentenceId(), "example", key, "diff", entityCounts[key] - exampleCounts[key]
return examples
def buildExamplesInner(self, sentenceGraph, goldGraph):
"""
Build one example for each token of the sentence
"""
if sentenceGraph.sentenceElement.get("origId") in self.skiplist:
print >> sys.stderr, "Skipping sentence", sentenceGraph.sentenceElement.get(
"origId")
return []
self.multiEdgeFeatureBuilder.setFeatureVector(resetCache=True)
self.triggerFeatureBuilder.initSentence(sentenceGraph)
undirected = self.nxMultiDiGraphToUndirected(
sentenceGraph.dependencyGraph)
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
# Get argument order
self.interactionLengths = self.getInteractionEdgeLengths(
sentenceGraph, paths)
self.interactionLengths = self.interactionLengths.values()
self.interactionLengths.sort(compareInteractionPrecedence)
# Map tokens to entities
tokenByOffset = {}
for i in range(len(sentenceGraph.tokens)):
token = sentenceGraph.tokens[i]
if goldGraph != None:
goldToken = goldGraph.tokens[i]
assert token.get("id") == goldToken.get("id") and token.get(
"charOffset") == goldToken.get("charOffset")
tokenByOffset[token.get("charOffset")] = token.get("id")
# Map gold entities to their head offsets
goldEntitiesByOffset = {}
for token in sentenceGraph.tokens:
goldEntitiesByOffset[token.get("charOffset")] = []
entityToGold = {}
for entity in sentenceGraph.entities:
entityToGold[entity] = []
if goldGraph != None:
for entity in goldGraph.entities:
offset = entity.get("headOffset")
assert offset != None
goldEntitiesByOffset[offset].append(entity)
# Map predicted entities to gold entities
for entity in sentenceGraph.entities:
eType = entity.get("type")
eOffset = entity.get("headOffset")
for goldEntity in goldEntitiesByOffset[eOffset]:
if goldEntity.get("type") == eType:
entityToGold[entity].append(goldEntity)
# Map entities to interactions
#interactionsByEntityId = {}
#for entity in sentenceGraph.entities:
# interactionsByEntityId[entity.get("id")] = []
# Map tokens to interactions
interactionsByToken = {}
for token in sentenceGraph.tokens:
interactionsByToken[token] = []
for interactionTuple in self.interactionLengths:
interaction = interactionTuple[0]
if interaction.get("type") == "neg":
continue
e1Id = interaction.get("e1")
token = sentenceGraph.entityHeadTokenByEntity[
sentenceGraph.entitiesById[e1Id]]
interactionsByToken[token].append(interaction)
examples = []
exampleIndex = 0
self.tokenFeatures = {}
#namedEntityNorStrings = set()
namedEntityHeadTokens = []
if not "names" in self.styles:
namedEntityCount = 0
for entity in sentenceGraph.entities:
if entity.get(
"isName"
) == "True": # known data which can be used for features
namedEntityCount += 1
#namedEntityNorStrings.add( entity.get("text").replace("-","").replace("/","").replace(",","").replace("\\","").replace(" ","").lower() )
namedEntityCountFeature = "nameCount_" + str(namedEntityCount)
#if namedEntityCount == 0: # no names, no need for triggers
# return []
if "pos_pairs" in self.styles:
namedEntityHeadTokens = self.getNamedEntityHeadTokens(
sentenceGraph)
#neFeatures = {} # F: 69.35 -> 69.14
#for norString in namedEntityNorStrings:
# neFeatures[self.featureSet.getId("norNE_" + norString)] = 1
bagOfWords = {}
for token in sentenceGraph.tokens:
text = "bow_" + token.get("text")
if not bagOfWords.has_key(text):
bagOfWords[text] = 0
bagOfWords[text] += 1
if sentenceGraph.tokenIsName[token]:
text = "ne_" + text
if not bagOfWords.has_key(text):
bagOfWords[text] = 0
bagOfWords[text] += 1
bowFeatures = {}
for k, v in bagOfWords.iteritems():
bowFeatures[self.featureSet.getId(k)] = v
self.inEdgesByToken = {}
self.outEdgesByToken = {}
self.edgeSetByToken = {}
for token in sentenceGraph.tokens:
inEdges = sentenceGraph.dependencyGraph.in_edges(token, data=True)
fixedInEdges = []
for edge in inEdges:
fixedInEdges.append((edge[0], edge[1], edge[2]["element"]))
inEdges = fixedInEdges
inEdges.sort(compareDependencyEdgesById)
self.inEdgesByToken[token] = inEdges
outEdges = sentenceGraph.dependencyGraph.out_edges(token,
data=True)
fixedOutEdges = []
for edge in outEdges:
fixedOutEdges.append((edge[0], edge[1], edge[2]["element"]))
outEdges = fixedOutEdges
outEdges.sort(compareDependencyEdgesById)
self.outEdgesByToken[token] = outEdges
self.edgeSetByToken[token] = set(inEdges + outEdges)
for i in range(len(sentenceGraph.tokens)):
token = sentenceGraph.tokens[i]
# Recognize only non-named entities (i.e. interaction words)
if sentenceGraph.tokenIsName[
token] and not "names" in self.styles and not "all_tokens" in self.styles:
continue
# CLASS
#if len(sentenceGraph.tokenIsEntityHead[token]) > 0:
# category = self.classSet.getId(self.getMergedEntityType(sentenceGraph.tokenIsEntityHead[token]))
#else:
# category = 1
offset = token.get("charOffset")
if len(goldEntitiesByOffset[offset]) > 0:
category = self.classSet.getId(
self.getMergedEntityType(goldEntitiesByOffset[offset]))
else:
category = 1
tokenText = token.get("text").lower()
if "stem_gazetteer" in self.styles:
tokenText = PorterStemmer.stem(tokenText)
if ("exclude_gazetteer" in self.styles
) and self.gazetteer and tokenText not in self.gazetteer:
features = {}
features[self.featureSet.getId("exclude_gazetteer")] = 1
extra = {
"xtype": "token",
"t": token.get("id"),
"excluded": "True"
}
examples.append(
(sentenceGraph.getSentenceId() + ".x" + str(exampleIndex),
category, features, extra))
exampleIndex += 1
continue
# FEATURES
features = {}
self.features = features
if not "names" in self.styles:
features[self.featureSet.getId(namedEntityCountFeature)] = 1
#for k,v in bagOfWords.iteritems():
# features[self.featureSet.getId(k)] = v
# pre-calculate bow _features_
features.update(bowFeatures)
#features.update(neFeatures)
# for j in range(len(sentenceGraph.tokens)):
# text = "bow_" + sentenceGraph.tokens[j].get("text")
# if j < i:
# features[self.featureSet.getId("bf_" + text)] = 1
# elif j > i:
# features[self.featureSet.getId("af_" + text)] = 1
# Main features
text = token.get("text")
features[self.featureSet.getId("txt_" + text)] = 1
features[self.featureSet.getId("POS_" + token.get("POS"))] = 1
stem = PorterStemmer.stem(text)
features[self.featureSet.getId("stem_" + stem)] = 1
features[self.featureSet.getId("nonstem_" + text[len(stem):])] = 1
# Normalized versions of the string (if same as non-normalized, overlap without effect)
normalizedText = text.replace("-", "").replace("/", "").replace(
",", "").replace("\\", "").replace(" ", "").lower()
if normalizedText == "bound": # should be for all irregular verbs
normalizedText = "bind"
features[self.featureSet.getId("txt_" + normalizedText)] = 1
norStem = PorterStemmer.stem(normalizedText)
features[self.featureSet.getId("stem_" + norStem)] = 1
features[self.featureSet.getId("nonstem_" +
normalizedText[len(norStem):])] = 1
if "gazetteer_features_maintoken" in self.styles:
tokTxtLower = text.lower()
if "stem_gazetteer" in self.styles:
tokTxtLower = PorterStemmer.stem(tokTxtLower)
if self.gazetteer and tokTxtLower in self.gazetteer:
for label, weight in self.gazetteer[tokTxtLower].items():
features[self.featureSet.getId(
"gaz_knownLabel_" +
label)] = weight # 1 performs slightly worse
# Linear order features
#for index in [-3,-2,-1,1,2,3,4,5]: # 69.35 -> 68.97
for index in [-3, -2, -1, 1, 2, 3]:
if i + index > 0 and i + index < len(sentenceGraph.tokens):
self.buildLinearOrderFeatures(sentenceGraph, i + index,
str(index), features)
# Content
if i > 0 and text[0].isalpha() and text[0].isupper():
features[self.featureSet.getId("upper_case_start")] = 1
for j in range(len(text)):
if j > 0 and text[j].isalpha() and text[j].isupper():
features[self.featureSet.getId("upper_case_middle")] = 1
# numbers and special characters
if text[j].isdigit():
features[self.featureSet.getId("has_digits")] = 1
if j > 0 and text[j - 1] == "-":
features[self.featureSet.getId(
"has_hyphenated_digit")] = 1
elif text[j] == "-":
features[self.featureSet.getId("has_hyphen")] = 1
elif text[j] == "/":
features[self.featureSet.getId("has_fslash")] = 1
elif text[j] == "\\":
features[self.featureSet.getId("has_bslash")] = 1
# duplets
if j > 0:
features[self.featureSet.getId("dt_" +
text[j - 1:j +
1].lower())] = 1
# triplets
if j > 1:
features[self.featureSet.getId("tt_" +
text[j - 2:j +
1].lower())] = 1
# Attached edges (Hanging in and out edges)
t1InEdges = self.inEdgesByToken[token]
for edge in t1InEdges:
edgeType = edge[2].get("type")
features[self.featureSet.getId("t1HIn_" + edgeType)] = 1
features[self.featureSet.getId("t1HIn_" +
edge[0].get("POS"))] = 1
features[self.featureSet.getId("t1HIn_" + edgeType + "_" +
edge[0].get("POS"))] = 1
tokenText = sentenceGraph.getTokenText(edge[0])
features[self.featureSet.getId("t1HIn_" + tokenText)] = 1
features[self.featureSet.getId("t1HIn_" + edgeType + "_" +
tokenText)] = 1
t1OutEdges = self.outEdgesByToken[token]
for edge in t1OutEdges:
edgeType = edge[2].get("type")
features[self.featureSet.getId("t1HOut_" + edgeType)] = 1
features[self.featureSet.getId("t1HOut_" +
edge[1].get("POS"))] = 1
features[self.featureSet.getId("t1HOut_" + edgeType + "_" +
edge[1].get("POS"))] = 1
tokenText = sentenceGraph.getTokenText(edge[1])
features[self.featureSet.getId("t1HOut_" + tokenText)] = 1
features[self.featureSet.getId("t1HOut_" + edgeType + "_" +
tokenText)] = 1
extra = {"xtype": "token", "t": token.get("id")}
examples.append(
(sentenceGraph.getSentenceId() + ".x" + str(exampleIndex),
category, features, extra))
exampleIndex += 1
# chains
self.buildChains(token, sentenceGraph, features)
if "pos_pairs" in self.styles:
self.buildPOSPairs(token, namedEntityHeadTokens, features)
self.buildPredictionFeatures(sentenceGraph, paths, token,
interactionsByToken[token])
return examples
def buildChains(self,
token,
sentenceGraph,
features,
depthLeft=3,
chain="",
visited=None):
if depthLeft == 0:
return
strDepthLeft = "dist_" + str(depthLeft)
if visited == None:
visited = set()
inEdges = self.inEdgesByToken[token]
outEdges = self.outEdgesByToken[token]
edgeSet = visited.union(self.edgeSetByToken[token])
for edge in inEdges:
if not edge in visited:
edgeType = edge[2].get("type")
features[self.featureSet.getId("dep_" + strDepthLeft +
edgeType)] = 1
nextToken = edge[0]
for tokenFeature, w in self.getTokenFeatures(
nextToken, sentenceGraph).iteritems():
features[self.featureSet.getId(strDepthLeft +
tokenFeature)] = w
# for entity in sentenceGraph.tokenIsEntityHead[nextToken]:
# if entity.get("isName") == "True":
# features[self.featureSet.getId("name_dist_"+strDepthLeft)] = 1
# features[self.featureSet.getId("name_dist_"+strDepthLeft+entity.get("type"))] = 1
# features[self.featureSet.getId("POS_dist_"+strDepthLeft+nextToken.get("POS"))] = 1
# tokenText = sentenceGraph.getTokenText(nextToken)
# features[self.featureSet.getId("text_dist_"+strDepthLeft+tokenText)] = 1
if sentenceGraph.tokenIsName[nextToken]:
features[self.featureSet.getId("name_chain_dist_" +
strDepthLeft + chain +
"-frw_" + edgeType)] = 1
features[self.featureSet.getId("chain_dist_" + strDepthLeft +
chain + "-frw_" + edgeType)] = 1
self.buildChains(nextToken, sentenceGraph, features,
depthLeft - 1, chain + "-frw_" + edgeType,
edgeSet)
for edge in outEdges:
if not edge in visited:
edgeType = edge[2].get("type")
features[self.featureSet.getId("dep_dist_" + strDepthLeft +
edgeType)] = 1
nextToken = edge[1]
for tokenFeature, w in self.getTokenFeatures(
nextToken, sentenceGraph).iteritems():
features[self.featureSet.getId(strDepthLeft +
tokenFeature)] = w
# for entity in sentenceGraph.tokenIsEntityHead[nextToken]:
# if entity.get("isName") == "True":
# features[self.featureSet.getId("name_dist_"+strDepthLeft)] = 1
# features[self.featureSet.getId("name_dist_"+strDepthLeft+entity.get("type"))] = 1
# features[self.featureSet.getId("POS_dist_"+strDepthLeft+nextToken.get("POS"))] = 1
# tokenText = sentenceGraph.getTokenText(nextToken)
# features[self.featureSet.getId("text_dist_"+strDepthLeft+tokenText)] = 1
if sentenceGraph.tokenIsName[nextToken]:
features[self.featureSet.getId("name_chain_dist_" +
strDepthLeft + chain +
"-rev_" + edgeType)] = 1
features[self.featureSet.getId("chain_dist_" + strDepthLeft +
chain + "-rev_" + edgeType)] = 1
self.buildChains(nextToken, sentenceGraph, features,
depthLeft - 1, chain + "-rev_" + edgeType,
edgeSet)
def getNamedEntityHeadTokens(self, sentenceGraph):
headTokens = []
for entity in sentenceGraph.entities:
if entity.get(
"isName"
) == "True": # known data which can be used for features
headTokens.append(
sentenceGraph.entityHeadTokenByEntity[entity])
return headTokens
def buildPOSPairs(self, token, namedEntityHeadTokens, features):
tokenPOS = token.get("POS")
assert tokenPOS != None
for headToken in namedEntityHeadTokens:
headPOS = headToken.get("POS")
features[self.featureSet.getId("POS_pair_NE_" + tokenPOS + "-" +
headPOS)] = 1
######################################################
# Unmerging-style features
######################################################
def buildPredictionFeatures(
self, sentenceGraph, paths, token,
interactions): #themeEntities, causeEntities=None):
# NOTE!!!! TODO
# add also features for arguments present, but not in this combination
self.buildInterArgumentBagOfWords(interactions, sentenceGraph)
if sentenceGraph.entitiesByToken.has_key(token):
for eventEntity in sentenceGraph.entitiesByToken[token]:
eventEntityType = eventEntity.get("type")
self.setFeature("rootType_" + eventEntity.get("type"), 1)
self.setFeature("predStrength" + eventEntityType,
self.getPredictionStrength(eventEntity))
self.triggerFeatureBuilder.setFeatureVector(self.features)
self.triggerFeatureBuilder.tag = "trg" + eventEntityType + "_"
self.triggerFeatureBuilder.buildFeatures(token)
self.triggerFeatureBuilder.tag = None
argThemeCount = 0
argCauseCount = 0
# Current example's edge combination
for i in range(len(interactions)):
arg = interactions[i]
if arg.get("type") == "Theme":
argThemeCount += 1
self.buildArgumentFeatures(sentenceGraph, paths, self.features,
token, arg, "argTheme")
self.buildArgumentFeatures(sentenceGraph, paths, self.features,
token, arg, "argTheme" + str(i))
else: # Cause
argCauseCount += 1
self.buildArgumentFeatures(sentenceGraph, paths, self.features,
token, arg, "argCause")
self.buildArgumentFeatures(sentenceGraph, paths, self.features,
token, arg, "argCause" + str(i))
self.setFeature("argCount", len(interactions))
self.setFeature("argCount_" + str(len(interactions)), 1)
self.setFeature("argThemeCount", argThemeCount)
self.setFeature("argThemeCount_" + str(argThemeCount), 1)
self.setFeature("argCauseCount", argCauseCount)
self.setFeature("argCauseCount_" + str(argCauseCount), 1)
self.triggerFeatureBuilder.tag = ""
self.triggerFeatureBuilder.setFeatureVector(None)
def buildArgumentFeatures(self, sentenceGraph, paths, features, eventToken,
arg, tag):
argEntity = sentenceGraph.entitiesById[arg.get("e2")]
argToken = sentenceGraph.entityHeadTokenByEntity[argEntity]
self.buildEdgeFeatures(sentenceGraph, paths, features, eventToken,
argToken, tag)
self.triggerFeatureBuilder.tag = tag + "trg_"
self.triggerFeatureBuilder.buildFeatures(argToken)
if argEntity.get("isName") == "True":
self.setFeature(tag + "Protein", 1)
else:
self.setFeature(tag + "Event", 1)
self.setFeature("nestingEvent", 1)
self.setFeature(tag + "_" + argEntity.get("type"), 1)
def buildEdgeFeatures(self, sentenceGraph, paths, features, eventToken,
argToken, tag):
#eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
#argToken = sentenceGraph.entityHeadTokenByEntity[argNode]
self.multiEdgeFeatureBuilder.tag = tag + "_"
self.multiEdgeFeatureBuilder.setFeatureVector(features, None, None,
False)
self.setFeature(tag + "_present", 1)
if eventToken != argToken and paths.has_key(
eventToken) and paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
edges = self.multiEdgeFeatureBuilder.getEdges(
sentenceGraph.dependencyGraph, path)
else:
path = [eventToken, argToken]
edges = None
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
#if not "disable_terminus_features" in self.styles:
# self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(
path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(
2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
4, path, edges, sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(
path, edges, sentenceGraph)
#self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None, None, None, False)
self.multiEdgeFeatureBuilder.tag = ""
def buildInterArgumentBagOfWords(self, arguments, sentenceGraph):
if len(arguments) < 2:
return
indexByToken = {}
for i in range(len(sentenceGraph.tokens)):
indexByToken[sentenceGraph.tokens[i]] = i
argTokenIndices = set()
for arg in arguments:
argEntity = sentenceGraph.entitiesById[arg.get("e2")]
argToken = sentenceGraph.entityHeadTokenByEntity[argEntity]
argTokenIndices.add(indexByToken[argToken])
minIndex = min(argTokenIndices)
maxIndex = max(argTokenIndices)
self.setFeature("argBoWRange", (maxIndex - minIndex))
self.setFeature("argBoWRange_" + str(maxIndex - minIndex), 1)
bow = set()
for i in range(minIndex + 1, maxIndex):
token = sentenceGraph.tokens[i]
if len(sentenceGraph.tokenIsEntityHead[token]
) == 0 and not sentenceGraph.tokenIsName[token]:
bow.add(token.get("text"))
bow = sorted(list(bow))
for word in bow:
self.setFeature("argBoW_" + word, 1)
if word in ["/", "-"]:
self.setFeature("argBoW_slashOrHyphen", 1)
if len(bow) == 1:
self.setFeature("argBoWonly_" + bow[0], 1)
if bow[0] in ["/", "-"]:
self.setFeature("argBoWonly_slashOrHyphen", 1)
class AsymmetricEventExampleBuilder(ExampleBuilder):
def __init__(self,
style=["typed", "directed"],
length=None,
types=[],
featureSet=None,
classSet=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert (classSet.getId("neg") == 1)
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
if style.find(",") != -1:
style = style.split(",")
self.styles = style
self.negFrac = None
self.posPairGaz = POSPairGazetteer()
for s in style:
if s.find("negFrac") != -1:
self.negFrac = float(s.split("_")[-1])
print >> sys.stderr, "Downsampling negatives to", self.negFrac
self.negRand = random.Random(15)
elif s.find("posPairGaz") != -1:
self.posPairGaz = POSPairGazetteer(
loadFrom=s.split("_", 1)[-1])
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
self.triggerFeatureBuilder = TriggerFeatureBuilder(self.featureSet)
if "graph_kernel" in self.styles:
from FeatureBuilders.GraphKernelFeatureBuilder import GraphKernelFeatureBuilder
self.graphKernelFeatureBuilder = GraphKernelFeatureBuilder(
self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
if "ontology" in self.styles:
self.multiEdgeFeatureBuilder.ontologyFeatureBuilder = BioInferOntologyFeatureBuilder(
self.featureSet)
if "nodalida" in self.styles:
self.nodalidaFeatureBuilder = NodalidaFeatureBuilder(
self.featureSet)
#IF LOCAL
if "bioinfer_limits" in self.styles:
self.bioinferOntologies = OntologyUtils.getBioInferTempOntology()
#self.bioinferOntologies = OntologyUtils.loadOntologies(OntologyUtils.g_bioInferFileName)
#ENDIF
self.pathLengths = length
assert (self.pathLengths == None)
self.types = types
if "random" in self.styles:
from FeatureBuilders.RandomFeatureBuilder import RandomFeatureBuilder
self.randomFeatureBuilder = RandomFeatureBuilder(self.featureSet)
#self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls, input, output, parse, tokenization, style, idFileTag=None):
classSet, featureSet = cls.getIdSets(idFileTag)
if style != None:
e = cls(style=style, classSet=classSet, featureSet=featureSet)
else:
e = cls(classSet=classSet, featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
if "printClassIds" in e.styles:
print >> sys.stderr, e.classSet.Ids
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(
sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def filterEdgesByType(self, edges, typesToInclude):
if len(typesToInclude) == 0:
return edges
edgesToKeep = []
for edge in edges:
if edge.get("type") in typesToInclude:
edgesToKeep.append(edge)
return edgesToKeep
def getCategoryNameFromTokens(self, sentenceGraph, t1, t2, directed=True):
types = set()
themeE1Types = set()
intEdges = []
if sentenceGraph.interactionGraph.has_edge(t1, t2):
intEdges = sentenceGraph.interactionGraph.get_edge_data(t1,
t2,
default={})
# NOTE: Only works if keys are ordered integers
for i in range(len(intEdges)):
types.add(intEdges[i]["element"].get("type"))
# if (not directed) and sentenceGraph.interactionGraph.has_edge(t2, t1):
# intEdgesReverse = sentenceGraph.interactionGraph.get_edge(t2, t1, default={})
# # NOTE: Only works if keys are ordered integers
# for i in range(len(intEdgesReverse)):
# intElement = intEdgesReverse[i]["element"]
# intType = intElement.get("type")
# types.add(intType)
# intEdges.extend(intEdgesReverse)
for i in range(len(intEdges)):
intElement = intEdges[i]["element"]
intType = intElement.get("type")
if intType == "Theme":
e1Entity = sentenceGraph.entitiesById[intElement.get("e1")]
themeE1Types.add(e1Entity.get("type"))
#types.add(intType)
if len(themeE1Types) != 0:
themeE1Types = list(themeE1Types)
themeE1Types.sort()
categoryName = ""
for name in themeE1Types:
if categoryName != "":
categoryName += "---"
categoryName += name
return categoryName
else:
types = list(types)
types.sort()
categoryName = ""
for name in types:
if categoryName != "":
categoryName += "---"
categoryName += name
if categoryName != "":
return categoryName
else:
return "neg"
def getCategoryName(self, sentenceGraph, e1, e2, directed=True):
interactions = sentenceGraph.getInteractions(e1, e2)
if not directed:
interactions.extend(sentenceGraph.getInteractions(e2, e1))
types = set()
for interaction in interactions:
types.add(interaction.attrib["type"])
types = list(types)
types.sort()
categoryName = ""
for name in types:
if categoryName != "":
categoryName += "---"
categoryName += name
if categoryName != "":
return categoryName
else:
return "neg"
def preProcessExamples(self, allExamples):
# Duplicates cannot be removed here, as they should only be removed from the training set. This is done
# in the classifier.
# if "no_duplicates" in self.styles:
# count = len(allExamples)
# print >> sys.stderr, " Removing duplicates,",
# allExamples = ExampleUtils.removeDuplicates(allExamples)
# print >> sys.stderr, "removed", count - len(allExamples)
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def isPotentialGeniaInteraction(self, e1, e2):
if e1.get("isName") == "True":
return False
else:
return True
#IF LOCAL
def getBioInferParentType(self, eType):
if eType == "Physical_entity" or OntologyUtils.hasParent(
eType, "Physical_entity", self.bioinferOntologies):
return "Physical"
elif eType == "Property_entity" or OntologyUtils.hasParent(
eType, "Property_entity", self.bioinferOntologies):
return "Property"
elif OntologyUtils.hasParent(eType, "Relationship",
self.bioinferOntologies):
return "Process"
else:
assert False, eType
# if self.bioinferOntologies["Entity"].has_key(eType):
# if OntologyUtils.hasParent(eType, "Physical_entity", self.bioinferOntologies):
# assert not OntologyUtils.hasParent(eType, "Property_entity", self.bioinferOntologies), eType
# return "Physical"
# else:
# assert OntologyUtils.hasParent(eType, "Property_entity", self.bioinferOntologies), eType
# return "Property"
#
# else:
# assert self.bioinferOntologies.has_key(eType), eType
# #assert OntologyUtils.hasParent(eType, "Process_entity", self.bioinferOntologies["Relationship"]), eType
# return "Process"
def isPotentialBioInferInteraction(self, e1, e2, categoryName):
e1Type = self.getBioInferParentType(e1.get("type"))
e2Type = self.getBioInferParentType(e2.get("type"))
if e1Type == "Process" or e1Type == "Property":
return True
elif e1Type == "Physical" and e2Type == "Physical":
return True
elif e1Type == "Physical" and e2Type == "Process": # hack
return True
else:
assert (
categoryName == "neg"
), categoryName + " category for " + e1Type + " and " + e2Type
return False
#ENDIF
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def buildExamples(self, sentenceGraph):
examples = []
exampleIndex = 0
clearGraph = sentenceGraph.getCleared()
#undirected = sentenceGraph.getUndirectedDependencyGraph()
undirected = self.nxMultiDiGraphToUndirected(
sentenceGraph.dependencyGraph)
##undirected = sentenceGraph.dependencyGraph.to_undirected()
###undirected = NX10.MultiGraph(sentenceGraph.dependencyGraph) This didn't work
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
self.triggerFeatureBuilder.initSentence(clearGraph)
# Generate examples based on interactions between entities or interactions between tokens
if "entities" in self.styles:
loopRange = len(sentenceGraph.entities)
else:
loopRange = len(sentenceGraph.tokens)
#for i in range(loopRange-1):
for i in range(loopRange): # allow self-interactions
#for j in range(i+1,loopRange):
for j in range(i, loopRange): # allow self-interactions
eI = None
eJ = None
if "entities" in self.styles:
eI = sentenceGraph.entities[i]
eJ = sentenceGraph.entities[j]
tI = sentenceGraph.entityHeadTokenByEntity[eI]
tJ = sentenceGraph.entityHeadTokenByEntity[eJ]
#if "no_ne_interactions" in self.styles and eI.get("isName") == "True" and eJ.get("isName") == "True":
# continue
if eI.get("type") == "neg" or eJ.get("type") == "neg":
continue
else:
tI = sentenceGraph.tokens[i]
tJ = sentenceGraph.tokens[j]
# # only consider paths between entities (NOTE! entities, not only named entities)
# if "headsOnly" in self.styles:
# if (len(sentenceGraph.tokenIsEntityHead[tI]) == 0) or (len(sentenceGraph.tokenIsEntityHead[tJ]) == 0):
# continue
if "directed" in self.styles:
# define forward
if "entities" in self.styles:
categoryName = self.getCategoryName(
sentenceGraph, eI, eJ, True)
else:
categoryName = self.getCategoryNameFromTokens(
sentenceGraph, tI, tJ, True)
self.exampleStats.beginExample(categoryName)
if self.negFrac == None or categoryName != "neg" or (
categoryName == "neg"
and self.negRand.random() < self.negFrac):
makeExample = True
if ("genia_limits" in self.styles
) and not self.isPotentialGeniaInteraction(eI, eJ):
makeExample = False
self.exampleStats.filter("genia_limits")
if self.posPairGaz.getNegFrac(
(tI.get("POS"), tJ.get("POS"))) == 1.0:
makeExample = False
self.exampleStats.filter("pos_pair")
if makeExample:
if not sentenceGraph.tokenIsName[tI]:
examples.append(
self.buildExample(tI, tJ, paths,
clearGraph, categoryName,
exampleIndex, eI, eJ))
exampleIndex += 1
else:
self.exampleStats.filter("genia_token_limits")
else:
self.exampleStats.filter("neg_frac")
self.exampleStats.endExample()
# define reverse
if "entities" in self.styles:
categoryName = self.getCategoryName(
sentenceGraph, eJ, eI, True)
else:
categoryName = self.getCategoryNameFromTokens(
sentenceGraph, tJ, tI, True)
self.exampleStats.beginExample(categoryName)
if self.negFrac == None or categoryName != "neg" or (
categoryName == "neg"
and self.negRand.random() < self.negFrac):
makeExample = True
if ("genia_limits" in self.styles
) and not self.isPotentialGeniaInteraction(eJ, eI):
makeExample = False
self.exampleStats.filter("genia_limits")
if ("bioinfer_limits" in self.styles
) and not self.isPotentialBioInferInteraction(
eJ, eI, categoryName):
makeExample = False
self.exampleStats.filter("bioinfer_limits")
if self.posPairGaz.getNegFrac(
(tJ.get("POS"), tI.get("POS"))) == 1.0:
makeExample = False
self.exampleStats.filter("pos_pair")
if makeExample:
if not sentenceGraph.tokenIsName[tJ]:
examples.append(
self.buildExample(tJ, tI, paths,
clearGraph, categoryName,
exampleIndex, eJ, eI))
exampleIndex += 1
else:
self.exampleStats.filter("genia_token_limits")
else:
self.exampleStats.filter("neg_frac")
self.exampleStats.endExample()
# else:
# if "entities" in self.styles:
# categoryName = self.getCategoryName(sentenceGraph, eI, eJ, False)
# else:
# categoryName = self.getCategoryNameFromTokens(sentenceGraph, tI, tJ, False)
# forwardExample = self.buildExample(tI, tJ, paths, clearGraph, categoryName, exampleIndex, eI, eJ)
# if not "graph_kernel" in self.styles:
# reverseExample = self.buildExample(tJ, tI, paths, clearGraph, categoryName, exampleIndex, eJ, eI)
# forwardExample[2].update(reverseExample[2])
# examples.append(forwardExample)
# exampleIndex += 1
return examples
def buildExample(self,
token1,
token2,
paths,
sentenceGraph,
categoryName,
exampleIndex,
entity1=None,
entity2=None):
# define features
features = {}
if True: #token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
if token1 != token2 and paths.has_key(
token1) and paths[token1].has_key(token2):
path = paths[token1][token2]
else:
path = [token1, token2]
assert (self.pathLengths == None)
if self.pathLengths == None or len(path) - 1 in self.pathLengths:
if not "no_trigger":
self.triggerFeatureBuilder.setFeatureVector(self.features)
self.triggerFeatureBuilder.tag = "trg_t1_"
self.triggerFeatureBuilder.buildFeatures(eventToken)
self.triggerFeatureBuilder.tag = "trg_t2_"
self.triggerFeatureBuilder.buildFeatures(eventToken)
# if not "no_ontology" in self.styles:
# self.ontologyFeatureBuilder.setFeatureVector(features)
# self.ontologyFeatureBuilder.buildOntologyFeaturesForPath(sentenceGraph, path)
# self.ontologyFeatureBuilder.setFeatureVector(None)
if "graph_kernel" in self.styles or not "no_dependency" in self.styles:
if token1 != token2 and paths.has_key(
token1) and paths[token1].has_key(token2):
edges = self.multiEdgeFeatureBuilder.getEdges(
sentenceGraph.dependencyGraph, path)
else:
edges = None
if "graph_kernel" in self.styles:
self.graphKernelFeatureBuilder.setFeatureVector(
features, entity1, entity2)
self.graphKernelFeatureBuilder.buildGraphKernelFeatures(
sentenceGraph, path, edges)
self.graphKernelFeatureBuilder.setFeatureVector(None)
if "entity_type" in self.styles:
features[self.featureSet.getId("e1_" +
entity1.attrib["type"])] = 1
features[self.featureSet.getId("e2_" +
entity2.attrib["type"])] = 1
features[self.featureSet.getId("distance_" +
str(len(path)))] = 1
if not "no_dependency" in self.styles:
if token1 == token2:
features[self.featureSet.getId("tokenSelfLoop")] = 1
self.multiEdgeFeatureBuilder.setFeatureVector(
features, entity1, entity2)
#self.multiEdgeFeatureBuilder.buildStructureFeatures(sentenceGraph, paths) # remove for fast
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(
sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
if not "disable_terminus_features" in self.styles:
self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(
path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(
path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(
2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
4, path, edges, sentenceGraph) # remove for fast
#self.buildEdgeCombinations(path, edges, sentenceGraph, features) # remove for fast
#if edges != None:
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[0], edges[0][1]+edges[1][0], "t1", sentenceGraph) # remove for fast
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[-1], edges[len(path)-1][len(path)-2]+edges[len(path)-2][len(path)-1], "t2", sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(
path, edges, sentenceGraph)
self.multiEdgeFeatureBuilder.buildSentenceFeatures(
sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None)
if "nodalida" in self.styles:
self.nodalidaFeatureBuilder.setFeatureVector(
features, entity1, entity2)
shortestPaths = self.nodalidaFeatureBuilder.buildShortestPaths(
sentenceGraph.dependencyGraph, path)
print shortestPaths
if len(shortestPaths) > 0:
self.nodalidaFeatureBuilder.buildNGrams(
shortestPaths, sentenceGraph)
self.nodalidaFeatureBuilder.setFeatureVector(None)
if not "no_linear" in self.styles:
self.tokenFeatureBuilder.setFeatureVector(features)
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == token1:
token1Index = i
if sentenceGraph.tokens[i] == token2:
token2Index = i
linearPreTag = "linfw_"
if token1Index > token2Index:
token1Index, token2Index = token2Index, token1Index
linearPreTag = "linrv_"
self.tokenFeatureBuilder.buildLinearOrderFeatures(
token1Index, sentenceGraph, 2, 2, preTag="linTok1")
self.tokenFeatureBuilder.buildLinearOrderFeatures(
token2Index, sentenceGraph, 2, 2, preTag="linTok2")
# Before, middle, after
# self.tokenFeatureBuilder.buildTokenGrams(0, token1Index-1, sentenceGraph, "bf")
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, "bw")
# self.tokenFeatureBuilder.buildTokenGrams(token2Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, "af")
# before-middle, middle, middle-after
# self.tokenFeatureBuilder.buildTokenGrams(0, token2Index-1, sentenceGraph, linearPreTag+"bf", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, linearPreTag+"bw", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, linearPreTag+"af", max=2)
self.tokenFeatureBuilder.setFeatureVector(None)
if "random" in self.styles:
self.randomFeatureBuilder.setFeatureVector(features)
self.randomFeatureBuilder.buildRandomFeatures(100, 0.01)
self.randomFeatureBuilder.setFeatureVector(None)
if "genia_limits" in self.styles:
e1Type = entity1.get("type")
e2Type = entity2.get("type")
assert (entity1.get("isName") == "False")
if entity2.get("isName") == "True":
features[self.featureSet.getId(
"GENIA_target_protein")] = 1
else:
features[self.featureSet.getId(
"GENIA_nested_event")] = 1
if e1Type.find(
"egulation"
) != -1: # leave r out to avoid problems with capitalization
if entity2.get("isName") == "True":
features[self.featureSet.getId(
"GENIA_regulation_of_protein")] = 1
else:
features[self.featureSet.getId(
"GENIA_regulation_of_event")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
path = [token1, token2]
self.triggerFeatureBuilder.tag = ""
self.triggerFeatureBuilder.setFeatureVector(None)
# define extra attributes
# if int(path[0].attrib["id"].split("_")[-1]) < int(path[-1].attrib["id"].split("_")[-1]):
# #extra = {"xtype":"edge","type":"i","t1":path[0],"t2":path[-1]}
# extra = {"xtype":"asym","type":"i","t1":path[0].get("id"),"t2":path[-1].get("id")}
# extra["deprev"] = False
# else:
# #extra = {"xtype":"edge","type":"i","t1":path[-1],"t2":path[0]}
# extra = {"xtype":"asym","type":"i","t1":path[-1].get("id"),"t2":path[0].get("id")}
# extra["deprev"] = True
extra = {
"xtype": "asym",
"type": "i",
"t1": token1.get("id"),
"t2": token2.get("id")
}
if entity1 != None:
#extra["e1"] = entity1
extra["e1"] = entity1.get("id")
if entity2 != None:
#extra["e2"] = entity2
extra["e2"] = entity2.get("id")
extra["categoryName"] = categoryName
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
if "binary" in self.styles:
if categoryName != "neg":
category = 1
else:
category = -1
categoryName = "i"
else:
category = self.classSet.getId(categoryName)
return (sentenceGraph.getSentenceId() + ".x" + str(exampleIndex),
category, features, extra)
class DirectEventExampleBuilder(ExampleBuilder):
def __init__(self,
style=["typed", "directed", "headsOnly"],
length=None,
types=[],
featureSet=None,
classSet=None,
gazetteer=None,
pathGazetteer=None,
negFrac=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert (classSet.getId("neg") == 1)
if gazetteer != None:
print >> sys.stderr, "Loading gazetteer from", gazetteer
self.gazetteer = Gazetteer.loadGztr(gazetteer)
else:
print >> sys.stderr, "No gazetteer loaded"
self.gazetteer = None
self.pathGazetteer = None
self.pathGazetteerDependencies = None
self.pathGazetteerPairs = None
if pathGazetteer != None:
print >> sys.stderr, "Loading path gazetteer from", pathGazetteer
self.pathGazetteer = PathGazetteer.load(pathGazetteer)
self.pathGazetteerDependencies = PathGazetteer.getDependencies(
self.pathGazetteer)
self.pathGazetteerPairs = PathGazetteer.getPairs(
self.pathGazetteer)
else:
print >> sys.stderr, "No path gazetteer loaded"
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
self.styles = style
self.negFrac = negFrac
print >> sys.stderr, "Downsampling negatives to", negFrac
self.negRand = random.Random()
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if True: #"noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.triggerFeatureBuilder = TriggerFeatureBuilder(self.featureSet)
#self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
#if "ontology" in self.styles:
# self.multiEdgeFeatureBuilder.ontologyFeatureBuilder = BioInferOntologyFeatureBuilder(self.featureSet)
self.pathLengths = length
assert (self.pathLengths == None)
self.types = types
self.eventsByOrigId = {}
self.headTokensByOrigId = {}
self.interSentenceEvents = set()
self.examplesByEventOrigId = {}
self.skippedByType = {}
self.skippedByTypeAndReason = {}
self.builtByType = {}
self.gazMatchCache = {}
#self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls,
input,
output,
parse,
tokenization,
style,
idFileTag=None,
gazetteer=None,
pathGazetteer=None,
negFrac=None):
classSet, featureSet = cls.getIdSets(idFileTag)
if style != None:
e = DirectEventExampleBuilder(style=style,
classSet=classSet,
featureSet=featureSet,
gazetteer=gazetteer,
pathGazetteer=pathGazetteer,
negFrac=negFrac)
else:
e = DirectEventExampleBuilder(classSet=classSet,
featureSet=featureSet,
gazetteer=gazetteer,
pathGazetteer=pathGazetteer,
negFrac=negFrac)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
e.printStats()
def getGazetteerMatch(self, string):
if string in self.gazMatchCache:
return self.gazMatchCache[string]
origString = string
if "stem_gazetteer" in self.styles:
string = PorterStemmer.stem(string)
if string in self.gazetteer:
self.gazMatchCache[origString] = string
return string
elif string.find("-") != -1:
replaced = string.replace("-", "")
else:
self.gazMatchCache[origString] = None
return None
if replaced in self.gazetteer:
self.gazMatchCache[origString] = replaced
return replaced
else:
splitted = string.rsplit("-", 1)[-1]
if splitted in self.gazetteer:
self.gazMatchCache[origString] = splitted
return splitted
else:
self.gazMatchCache[origString] = None
return None
def isInGazetteer(self, string):
return self.getGazetteerMatch(string) != None
def printStats(self):
eventsByType = {}
for event in self.eventsByOrigId.values():
eventsByType[event.get("type")] = eventsByType.get(
event.get("type"), 0) + 1
f = open("missed-events", "wt")
missedEvents = {}
for key in self.examplesByEventOrigId.keys():
if self.examplesByEventOrigId[key] == 0:
if not missedEvents.has_key(
self.eventsByOrigId[key].get("type")):
missedEvents[self.eventsByOrigId[key].get("type")] = []
missedEvents[self.eventsByOrigId[key].get("type")].append(key)
for key in sorted(missedEvents.keys()):
f.write(key + "\n")
for id in sorted(missedEvents[key]):
f.write(" " + id + " ")
if id in self.interSentenceEvents:
f.write("intersentence ")
text = self.headTokensByOrigId[id].get("text").lower()
if not self.isInGazetteer(text):
text = self.headTokensByOrigId[id].get("text").lower()
if "stem_gazetteer" in self.styles:
stemmed = PorterStemmer.stem(text)
f.write("not-in-gazetteer (" + text + " / " + stemmed +
")")
f.write("\n")
f.close()
print >> sys.stderr, "Example selection missed events (other, intersentence, non-gazetteer)"
for key in sorted(eventsByType.keys()):
inter = 0
other = 0
nongaz = 0
if missedEvents.has_key(key):
for id in missedEvents[key]:
tokText = self.headTokensByOrigId[id].get("text").lower()
if id in self.interSentenceEvents:
inter += 1
elif not self.isInGazetteer(tokText):
nongaz += 1
else:
other += 1
if inter == other == nongaz == 0:
print >> sys.stderr, " " + key + " (" + str(
eventsByType[key]) + "): missed none"
else:
print >> sys.stderr, " " + key + " (" + str(
eventsByType[key]) + "): " + str(other) + ", " + str(
inter) + ", " + str(nongaz)
print >> sys.stderr, "Example generation (total, built/skipped)"
for key in sorted(
list(set(self.skippedByType.keys() +
self.builtByType.keys()))):
string = " " + key + ": (" + str(
self.builtByType.get(key, 0) +
self.skippedByType.get(key, 0)) + ", " + str(
self.builtByType.get(key, 0)) + "/" + str(
self.skippedByType.get(key, 0)) + ") ["
for key2 in sorted(self.skippedByTypeAndReason[key].keys()):
string += key2 + ":" + str(
self.skippedByTypeAndReason[key][key2]) + " "
string += "]"
print >> sys.stderr, string
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(
sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
# def isPotentialGeniaInteraction(self, e1, e2):
# if e1.get("isName") == "True" and e2.get("isName") == "True":
# return False
# elif e1.get("isName") == "True" and e2.get("isName") == "False":
# return False
# else:
# return True
def getArgumentEntities(self, sentenceGraph, entityNode):
eId = entityNode.get("id")
assert (eId != None)
themeNodes = []
causeNodes = []
for edge in sentenceGraph.interactions:
if edge.get("e1") == eId:
edgeType = edge.get("type")
assert (edgeType in ["Theme", "Cause"]), edgeType
if edgeType == "Theme":
themeNodes.append(
sentenceGraph.entitiesById[edge.get("e2")])
elif edgeType == "Cause":
causeNodes.append(
sentenceGraph.entitiesById[edge.get("e2")])
return themeNodes, causeNodes
def makeGSEvents(self, sentenceGraph):
self.namedEntityHeadTokenIds = set()
self.gsEvents = {} # [token]->[event-type]->[1-n argument sets]
for token in sentenceGraph.tokens:
self.gsEvents[token] = {}
for entity in sentenceGraph.entities:
if entity.get("type") == "neg":
continue
elif entity.get("isName") == "True":
self.namedEntityHeadTokenIds.add(
sentenceGraph.entityHeadTokenByEntity[entity].get("id"))
continue
eId = entity.get("id")
eOrigId = entity.get("origId")
assert not self.eventsByOrigId.has_key(eOrigId)
self.eventsByOrigId[eOrigId] = entity
if not self.examplesByEventOrigId.has_key(eOrigId):
self.examplesByEventOrigId[eOrigId] = 0
if len(sentenceGraph.interSentenceInteractions) > 0:
for interaction in sentenceGraph.interSentenceInteractions:
if interaction.get("e1") == eId:
self.interSentenceEvents.add(eOrigId)
eType = entity.get("type")
arguments = set()
for interaction in sentenceGraph.interactions:
if interaction.get("e1") == eId:
e2 = sentenceGraph.entitiesById[interaction.get("e2")]
e2TokenId = sentenceGraph.entityHeadTokenByEntity[e2].get(
"id")
arguments.add((interaction.get("type"), e2TokenId))
#arguments.add( (interaction.get("type"), interaction.get("e2") ) )
arguments = tuple(sorted(list(arguments)))
eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
self.headTokensByOrigId[eOrigId] = eHeadToken
if not self.gsEvents[eHeadToken].has_key(eType):
self.gsEvents[eHeadToken][eType] = {}
if len(arguments) > 0:
if not self.gsEvents[eHeadToken][eType].has_key(arguments):
self.gsEvents[eHeadToken][eType][arguments] = []
self.gsEvents[eHeadToken][eType][arguments].append(eOrigId)
def getGSEventType(self, sentenceGraph, eHeadToken, themeTokens,
causeTokens):
#eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
#eType = entity.get("type")
if len(self.gsEvents[eHeadToken]) == 0:
return "neg", []
argumentSet = set()
for themeNode in themeTokens:
if themeNode != None:
argumentSet.add(("Theme", themeNode.get("id")))
for causeNode in causeTokens:
if causeNode != None:
argumentSet.add(("Cause", causeNode.get("id")))
argumentSet = tuple(sorted(list(argumentSet)))
gsTypes = set()
eventIds = []
for eventType in sorted(self.gsEvents[eHeadToken].keys()):
if argumentSet in self.gsEvents[eHeadToken][eventType].keys():
gsTypes.add(eventType)
eventIds.extend(
self.gsEvents[eHeadToken][eventType][argumentSet])
if len(gsTypes) == 0:
return "neg", eventIds
elif len(gsTypes) == 1:
return list(gsTypes)[0], eventIds
else:
gsTypes = sorted(list(gsTypes))
string = gsTypes[0]
for gsType in gsTypes[1:]:
string += "---" + gsType
return string, eventIds
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def buildExamples(self, sentenceGraph):
self.makeGSEvents(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(resetCache=True)
self.triggerFeatureBuilder.initSentence(sentenceGraph)
examples = []
exampleIndex = 0
#undirected = sentenceGraph.dependencyGraph.to_undirected()
undirected = self.nxMultiDiGraphToUndirected(
sentenceGraph.dependencyGraph)
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
eventTokens = []
nameTokens = []
gazCategories = {None: {"neg": -1}}
#stems = {}
for token in sentenceGraph.tokens:
gazText = self.getGazetteerMatch(token.get("text").lower())
if gazText != None:
gazCategories[token] = self.gazetteer[gazText]
else:
gazCategories[token] = {"neg": -1}
if token.get("id") in self.namedEntityHeadTokenIds:
nameTokens.append(token)
elif gazText != None:
eventTokens.append(token)
allTokens = eventTokens + nameTokens
#if len(nameTokens) == 0: # there can be no events in this sentence
# self.gsEvents = None
# return []
for token in eventTokens:
#gazCategories = self.gazetteer[token.get("text").lower()]
#print token.get("text").lower(), gazCategories
#multiargument = False
potentialRegulation = False
potentialBinding = False
for key in gazCategories[token].keys():
if key in [
"Regulation", "Positive_regulation",
"Negative_regulation"
]:
#multiargument = True
potentialRegulation = True
break
for key in gazCategories[token].keys():
if key in ["Binding"]:
#multiargument = True
potentialBinding = True
break
if potentialRegulation:
combinations = combine.combine(allTokens, allTokens + [None])
else:
combinations = []
for t2 in nameTokens: #allTokens:
combinations.append((t2, None))
if potentialBinding:
for i in range(len(nameTokens) - 1):
for j in range(i + 1, len(nameTokens)):
combinations.append(
((nameTokens[i], nameTokens[j]), None))
for combination in combinations:
theme2Binding = False
if type(combination[0]) == types.ListType or type(
combination[0]) == types.TupleType:
theme2Binding = True
categoryName, eventIds = self.getGSEventType(
sentenceGraph, token, combination[0], [combination[1]])
else:
categoryName, eventIds = self.getGSEventType(
sentenceGraph, token, [combination[0]],
[combination[1]])
for id in eventIds:
self.examplesByEventOrigId[id] += 1
skip = False
s = self.skippedByTypeAndReason
if not s.has_key(categoryName):
s[categoryName] = {}
if gazCategories[token].get("neg", -1) > 0.99:
pass
if combination[0] == combination[1]:
pass #skip = True
if combination[0] == token or combination[1] == token:
if theme2Binding or gazCategories[combination[0]].get(
"Positive_regulation", -1) < 0:
skip = True
s[categoryName]["duparg"] = s[categoryName].get(
"duparg", 0) + 1
if combination[0] == None and combination[1] == None:
skip = True
s[categoryName]["noncmb"] = s[categoryName].get(
"noncmb", 0) + 1
validCat = self.isValidEvent(paths, sentenceGraph, token,
combination)
if validCat != "OK": #not self.isValidEvent(paths, sentenceGraph, token, combination):
skip = True
#s[categoryName]["valid"] = s[categoryName].get("valid", 0) + 1
s[categoryName][validCat] = s[categoryName].get(
validCat, 0) + 1
if len(nameTokens) == 0:
skip = True
s[categoryName]["non"] = s[categoryName].get("non", 0) + 1
if theme2Binding:
if gazCategories[combination[0][0]].get(
"neg",
-1) > 0.99 or gazCategories[combination[0][1]].get(
"neg", -1) > 0.99:
skip = True
s[categoryName]["gazarg"] = s[categoryName].get(
"gazarg", 0) + 1
else:
if gazCategories[combination[0]].get(
"neg",
-1) > 0.99 or gazCategories[combination[1]].get(
"neg", -1) > 0.99:
skip = True
s[categoryName]["gazarg"] = s[categoryName].get(
"gazarg", 0) + 1
if (skip and self.negFrac
== None) or (skip and self.negFrac != None
and categoryName == "neg"):
self.skippedByType[categoryName] = self.skippedByType.get(
categoryName, 0) + 1
else:
if self.negFrac == None or categoryName != "neg" or (
categoryName == "neg"
and self.negRand.random() < self.negFrac):
self.builtByType[categoryName] = self.builtByType.get(
categoryName, 0) + 1
if theme2Binding:
newExample = self.buildExample(
exampleIndex, sentenceGraph, paths, token,
combination[0], [combination[1]])
else:
newExample = self.buildExample(
exampleIndex, sentenceGraph, paths, token,
[combination[0]], [combination[1]])
if len(eventIds) > 0:
newExample[3]["numEv"] = str(len(eventIds))
examples.append(newExample)
exampleIndex += 1
self.gsEvents = None
return examples
def isValidEvent(self, paths, sentenceGraph, eventToken, argTokens):
# This one lets through Positive_regulations that are
# excluded from the duparg-rule
oneTokenEvent = True
for argToken in argTokens:
if argToken != None and eventToken != argToken:
oneTokenEvent = False
break
if oneTokenEvent:
return "OK" #True
if not paths.has_key(eventToken):
return "nopaths" #False
newArgTokens = []
for argToken in argTokens:
if type(argToken) == types.ListType or type(
argToken) == types.TupleType:
newArgTokens.extend(argToken)
else:
newArgTokens.append(argToken)
argTokens = newArgTokens
oneArgValid = True
if False:
oneArgValid = False
for argToken in argTokens:
if argToken == None:
continue
if paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
else:
#print argToken, argToken.get("text")
#return False
continue
depPaths = self.multiEdgeFeatureBuilder.getEdgeCombinations(
sentenceGraph.dependencyGraph, path)
validArg = False
for p in depPaths:
if p in self.pathGazetteer and self.pathGazetteer[p][0] > 0:
validArg = True
break
if validArg:
oneArgValid = True
# The first and last dependency of a path
if False:
oneEdgeValid = False
for argToken in argTokens:
if argToken == None:
continue
if paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
else:
#print argToken, argToken.get("text")
#return False
continue
depPaths = self.multiEdgeFeatureBuilder.getEdgeCombinations(
sentenceGraph.dependencyGraph, path)
validArg = False
for p in depPaths:
p = p.replace("<", "")
p = p.replace(">", "")
p = p.split(".")
pair = (p[0], p[-1])
if pair in self.pathGazetteerPairs:
validArg = True
break
if validArg:
oneEdgeValid = True
break
if not oneEdgeValid:
return "pair"
# Event must not have unseen dependencies in any of its paths
if False:
for argToken in argTokens:
if argToken == None:
continue
if paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
else:
continue
deps = self.multiEdgeFeatureBuilder.getEdgeSet(
sentenceGraph.dependencyGraph, path)
for d in deps:
if d[2].get("type") not in self.pathGazetteerDependencies:
#print "Unk", d[2].get("type")
return "unkdep"
# validArg = True
# for p in depPaths:
# if p in self.pathGazetteer and self.pathGazetteer[p][0] == 0:
# validArg = False
# break
# if not validArg:
# return False
if not oneArgValid:
return "novalidarg" #False
return "OK" #True
def setGazetteerFeatures(self, token, tag):
gazText = self.getGazetteerMatch(token.get("text").lower())
if gazText != None:
gazCategories = self.gazetteer[gazText]
for k, v in gazCategories.iteritems():
self.setFeature(tag + "gaz_event_value_" + k, v)
self.setFeature(tag + "gaz_event_" + k, 1)
if k.find("egulation") != -1:
self.setFeature(tag + "potReg", 1)
else:
self.setFeature(tag + "notInGaz", 1)
def buildExample(self,
exampleIndex,
sentenceGraph,
paths,
eventToken,
themeTokens,
causeTokens=None):
features = {}
self.features = features
categoryName, eventIds = self.getGSEventType(sentenceGraph, eventToken,
themeTokens, causeTokens)
category = self.classSet.getId(categoryName)
potentialRegulation = False
eventTokenText = eventToken.get("text").lower()
gazText = self.getGazetteerMatch(eventTokenText)
gazCategories = self.gazetteer[gazText]
for k, v in gazCategories.iteritems():
if k.find("egulation") != -1:
potentialRegulation = True
self.setGazetteerFeatures(eventToken, "")
self.triggerFeatureBuilder.setFeatureVector(self.features)
self.triggerFeatureBuilder.tag = "trg_"
self.triggerFeatureBuilder.buildFeatures(eventToken)
themeEntities = []
hasTheme = False
if len(themeTokens) > 1:
self.setFeature("multiTheme", 1)
potentialRegulation = False
for themeToken in themeTokens:
if themeToken != None:
hasTheme = True
self.setGazetteerFeatures(themeToken, "theme_")
self.buildArgumentFeatures(sentenceGraph, paths, features,
eventToken, themeToken, "theme_")
self.triggerFeatureBuilder.tag = "ttrg_"
self.triggerFeatureBuilder.buildFeatures(themeToken)
themeEntity = None
if sentenceGraph.entitiesByToken.has_key(themeToken):
for themeEntity in sentenceGraph.entitiesByToken[
themeToken]:
if themeEntity.get("isName") == "True":
self.setFeature("themeProtein", 1)
if potentialRegulation:
self.setFeature("regulationThemeProtein", 1)
themeEntities.append(themeEntity)
break
if not features.has_key("themeProtein"):
self.setFeature("themeEvent", 1)
self.setFeature("nestingEvent", 1)
if potentialRegulation:
self.setFeature("regulationThemeEvent", 1)
if hasTheme:
self.setFeature("noTheme", 1)
causeEntities = []
hasCause = False
for causeToken in causeTokens:
if causeToken != None:
hasCause = True
self.setGazetteerFeatures(causeToken, "cause_")
self.buildArgumentFeatures(sentenceGraph, paths, features,
eventToken, causeToken, "cause_")
self.triggerFeatureBuilder.tag = "ctrg_"
self.triggerFeatureBuilder.buildFeatures(causeToken)
causeEntity = None
if sentenceGraph.entitiesByToken.has_key(causeToken):
for causeEntity in sentenceGraph.entitiesByToken[
causeToken]:
if causeEntity.get("isName") == "True":
self.setFeature("causeProtein", 1)
if potentialRegulation:
self.setFeature("regulationCauseProtein", 1)
causeEntities.append(causeEntity)
break
if not features.has_key("causeProtein"):
self.setFeature("causeEvent", 1)
self.setFeature("nestingEvent", 1)
if potentialRegulation:
self.setFeature("regulationCauseEvent", 1)
if not hasCause:
self.setFeature("noCause", 1)
self.triggerFeatureBuilder.tag = ""
self.triggerFeatureBuilder.setFeatureVector(None)
# Common features
# if e1Type.find("egulation") != -1: # leave r out to avoid problems with capitalization
# if entity2.get("isName") == "True":
# features[self.featureSet.getId("GENIA_regulation_of_protein")] = 1
# else:
# features[self.featureSet.getId("GENIA_regulation_of_event")] = 1
# define extra attributes
extra = {"xtype": "event", "type": categoryName}
extra["et"] = eventToken.get("id")
if len(eventIds) > 0:
eventIds.sort()
extra["eids"] = ""
for eventId in eventIds:
extra["eids"] += str(eventId) + ","
extra["eids"] = extra["eids"][:-1]
for themeToken in themeTokens:
if themeToken != None:
if extra.has_key("tt"):
extra["tt"] = extra["tt"] + "," + themeToken.get("id")
else:
extra["tt"] = themeToken.get("id")
for themeEntity in themeEntities:
if extra.has_key("t"):
extra["t"] = extra["t"] + "," + themeEntity.get("id")
else:
extra["t"] = themeEntity.get("id")
for causeToken in causeTokens:
if causeToken != None:
extra["ct"] = causeTokens[0].get("id")
if len(causeEntities) > 0:
extra["c"] = causeEntities[0].get("id")
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
#assert (category == 1 or category == -1)
self.features = None
return (sentenceGraph.getSentenceId() + ".x" + str(exampleIndex),
category, features, extra)
def buildArgumentFeatures(self, sentenceGraph, paths, features, eventToken,
argToken, tag):
#eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
#argToken = sentenceGraph.entityHeadTokenByEntity[argNode]
self.multiEdgeFeatureBuilder.tag = tag
self.multiEdgeFeatureBuilder.setFeatureVector(features, None, None,
False)
if eventToken != argToken and paths.has_key(
eventToken) and paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
edges = self.multiEdgeFeatureBuilder.getEdges(
sentenceGraph.dependencyGraph, path)
else:
path = [eventToken, argToken]
edges = None
# if not "disable_entity_features" in self.styles:
# # doesn't improve beyond 52.32
# self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
# # buildPathLengthFeatures 52.32 -> 51-51
# self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
# if not "disable_terminus_features" in self.styles:
# # didn't improve from 52.32
# self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
# 50.74 -> 52.32
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(
path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
# ngrams alone - 50.74
self.multiEdgeFeatureBuilder.buildPathGrams(
2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
4, path, edges, sentenceGraph) # remove for fast
# disabling length 4 drops performance
# if not "disable_path_edge_features" in self.styles:
# self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(path, edges, sentenceGraph)
# self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
# buildSentenceFeatures seems to decrease performance by 8 %-points
self.multiEdgeFeatureBuilder.setFeatureVector(None, None, None, False)
self.multiEdgeFeatureBuilder.tag = ""
class UnmergedEdgeExampleBuilder(ExampleBuilder):
def __init__(self,
style=["typed", "directed", "headsOnly"],
length=None,
types=[],
featureSet=None,
classSet=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert (classSet.getId("neg") == 1)
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
self.styles = style
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
self.pathLengths = length
assert (self.pathLengths == None)
self.types = types
if "random" in self.styles:
from FeatureBuilders.RandomFeatureBuilder import RandomFeatureBuilder
self.randomFeatureBuilder = RandomFeatureBuilder(self.featureSet)
#self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls, input, output, parse, tokenization, style, idFileTag=None):
classSet, featureSet = cls.getIdSets(idFileTag)
if style == None:
e = UnmergedEdgeExampleBuilder(classSet=classSet,
featureSet=featureSet)
else:
e = UnmergedEdgeExampleBuilder(style=style,
classSet=classSet,
featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
print e.classSet.Ids
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(
sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def filterEdgesByType(self, edges, typesToInclude):
if len(typesToInclude) == 0:
return edges
edgesToKeep = []
for edge in edges:
if edge.get("type") in typesToInclude:
edgesToKeep.append(edge)
return edgesToKeep
def getCategoryName(self, sentenceGraph, e1, e2, directed=True):
# Dummies are potential entities that do not exist in the
# training data. If both entities of an interaction are dummies
# it can't exist in the training data and is therefore a negative
if e1[2] or e2[2]:
return "neg"
e1 = e1[0]
e2 = e2[0]
interactions = sentenceGraph.getInteractions(e1, e2)
if not directed:
interactions.extend(sentenceGraph.getInteractions(e2, e1))
types = set()
for interaction in interactions:
types.add(interaction.attrib["type"])
types = list(types)
types.sort()
categoryName = ""
for name in types:
if categoryName != "":
categoryName += "---"
categoryName += name
if categoryName != "":
return categoryName
else:
return "neg"
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def isPotentialGeniaInteraction(self, e1, e2):
if e1.get("isName") == "True":
return False
else:
return True
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def getInteractionEdgeLengths(self, sentenceGraph, paths):
"""
Return dependency and linear length of all interaction edges
(measured between the two tokens).
"""
interactionLengths = {}
for interaction in sentenceGraph.interactions:
# Calculated interaction edge dep and lin length
e1 = sentenceGraph.entitiesById[interaction.get("e1")]
e2 = sentenceGraph.entitiesById[interaction.get("e2")]
t1 = sentenceGraph.entityHeadTokenByEntity[e1]
t2 = sentenceGraph.entityHeadTokenByEntity[e2]
# Get dep path length
if t1 != t2 and paths.has_key(t1) and paths[t1].has_key(t2):
pathLength = len(paths[t1][t2])
else: # no dependencyPath
pathLength = 999999 # more than any real path
# Linear distance
t1Pos = -1
t2Pos = -1
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == t1:
t1Pos = i
if t2Pos != -1:
break
if sentenceGraph.tokens[i] == t2:
t2Pos = i
if t1Pos != -1:
break
linLength = abs(t1Pos - t2Pos)
interactionLengths[interaction] = (pathLength, linLength)
return interactionLengths
def getPrecedenceLevels(self, sentenceGraph, paths):
"""
Get overlapping entity precedence
"""
interactionLengths = self.getInteractionEdgeLengths(
sentenceGraph, paths)
interactionsByEntity = {} # Convenience mapping
entityPrecedenceValues = {}
for entity in sentenceGraph.entities:
interactionsByEntity[entity] = []
eId = entity.get("id")
# Add access to interactions
argDepDist = 0 # Sum of lengths of shortest paths
argLinDist = 0 # Sum of linear distances
for interaction in sentenceGraph.interactions:
if interaction.get(
"e1"
) == eId: # An argument of the entity defined by the node
interactionsByEntity[entity].append(interaction)
argDepDist += interactionLengths[interaction][0]
argLinDist += interactionLengths[interaction][1]
# Store precedence counts (num args, sum of dep lengths, sum of lin lengths)
entityPrecedenceValues[entity] = (len(interactionsByEntity),
argDepDist, argLinDist, entity)
# Determine level of entity from precedence counts
levelByEntity = {} # slot number
#levelByInteraction = {} # slot number of parent node
# There is one slot group per token, per type
for token in sentenceGraph.tokens: # per token
entitiesByType = {}
for entity in sentenceGraph.tokenIsEntityHead[token]: # per type
if entity.get(
"isName") == "True": # Names can never have duplicates
assert not levelByEntity.has_key(entity)
levelByEntity[entity] = 0
continue
eType = entity.get("type")
if eType == "neg":
continue
if not entitiesByType.has_key(eType):
entitiesByType[eType] = []
entitiesByType[eType].append(entity)
for eType in sorted(entitiesByType.keys()):
# Slot ordering by precedence
sortedEntities = []
for entity in entitiesByType[eType]:
sortedEntities.append(entityPrecedenceValues[entity])
sortedEntities.sort(compareEntityPrecedence)
level = 0
for precedenceTuple in sortedEntities:
entity = precedenceTuple[3]
assert not levelByEntity.has_key(entity)
levelByEntity[entity] = level
# Interactions have the same slot as their parent entity
#for interaction in interactionsByEntity[entity]:
# assert not levelByInteraction.has_key(interaction)
# levelByInteraction[interaction] = level
level += 1
return levelByEntity #, levelByInteraction
def buildExamples(self, sentenceGraph):
examples = []
exampleIndex = 0
#undirected = sentenceGraph.getUndirectedDependencyGraph()
undirected = self.nxMultiDiGraphToUndirected(
sentenceGraph.dependencyGraph)
##undirected = sentenceGraph.dependencyGraph.to_undirected()
###undirected = NX10.MultiGraph(sentenceGraph.dependencyGraph) This didn't work
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
# Determine overlapping entity precedence
#levelByEntity, levelByInteraction = self.getPrecedenceLevels(sentenceGraph, paths)
levelByEntity = self.getPrecedenceLevels(sentenceGraph, paths)
entities = []
# There is one entity group for each token, for each type of entity
for token in sentenceGraph.tokens: # per token
entitiesByType = {}
for entity in sentenceGraph.tokenIsEntityHead[token]: # per type
if entity.get(
"isName") == "True": # Names can never have duplicates
entities.append((entity, 0, False))
continue
eType = entity.get("type")
if eType == "neg":
continue
if not entitiesByType.has_key(eType):
entitiesByType[eType] = []
entitiesByType[eType].append(entity)
# Create slot groups for tokens for which exists at least one entity
eTypes = sorted(entitiesByType.keys())
if len(eTypes) == 0:
continue
# Create slot groups and insert GS data there
for eType in eTypes:
# Use first entity of a type as the dummy entity for unfilled slots
dummyEntity = entitiesByType[eType][0]
# Define entity slots
entityGroup = [None, None, None, None]
#entityGroup = [None, None]
# Insert existing entities into slots
for entity in entitiesByType[eType]:
if levelByEntity.has_key(entity):
level = levelByEntity[entity]
if level < len(entityGroup):
entityGroup[level] = (entity, level, False)
# Create dummies for potential entities
for i in range(len(entityGroup)):
if entityGroup[i] == None:
entityGroup[i] = (dummyEntity, i, True)
# Put all slots into one potential entity list
#print entityGroup
for e in entityGroup:
entities.append(e)
# Generate examples based on interactions between entities
for i in range(len(entities) - 1):
for j in range(i + 1, len(entities)):
eI = entities[i][0]
eJ = entities[j][0]
tI = sentenceGraph.entityHeadTokenByEntity[eI]
tJ = sentenceGraph.entityHeadTokenByEntity[eJ]
# define forward example
categoryName = self.getCategoryName(sentenceGraph, entities[i],
entities[j], True)
if (not "genia_limits"
in self.styles) or self.isPotentialGeniaInteraction(
eI, eJ):
examples.append(
self.buildExample(tI, tJ, paths, sentenceGraph,
categoryName, exampleIndex,
entities[i], entities[j]))
exampleIndex += 1
# define reverse
categoryName = self.getCategoryName(sentenceGraph, entities[j],
entities[i], True)
if (not "genia_limits"
in self.styles) or self.isPotentialGeniaInteraction(
eJ, eI):
examples.append(
self.buildExample(tJ, tI, paths, sentenceGraph,
categoryName, exampleIndex,
entities[j], entities[i]))
exampleIndex += 1
return examples
def buildExample(self,
token1,
token2,
paths,
sentenceGraph,
categoryName,
exampleIndex,
e1=None,
e2=None):
entity1 = e1[0]
entity2 = e2[0]
# define features
features = {}
features[self.featureSet.getId("gov_level")] = e1[1]
features[self.featureSet.getId("gov_level_" + str(e1[1]))] = 1
features[self.featureSet.getId("dep_level")] = e2[1]
features[self.featureSet.getId("dep_level_" + str(e2[1]))] = 1
features[self.featureSet.getId("level_pair_" + str(e1[1]) + "_" +
str(e2[1]))] = 1
if True: #token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
if token1 != token2 and paths.has_key(
token1) and paths[token1].has_key(token2):
path = paths[token1][token2]
else:
path = [token1, token2]
assert (self.pathLengths == None)
if self.pathLengths == None or len(path) - 1 in self.pathLengths:
if not "no_dependency" in self.styles:
if token1 != token2 and paths.has_key(
token1) and paths[token1].has_key(token2):
edges = self.multiEdgeFeatureBuilder.getEdges(
sentenceGraph.dependencyGraph, path)
else:
edges = None
if "entity_type" in self.styles:
features[self.featureSet.getId("e1_" +
entity1.attrib["type"])] = 1
features[self.featureSet.getId("e2_" +
entity2.attrib["type"])] = 1
features[self.featureSet.getId("distance_" +
str(len(path)))] = 1
if not "no_dependency" in self.styles:
self.multiEdgeFeatureBuilder.setFeatureVector(
features, entity1, entity2)
#self.multiEdgeFeatureBuilder.buildStructureFeatures(sentenceGraph, paths) # remove for fast
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(
sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
if not "disable_terminus_features" in self.styles:
self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(
path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(
path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(
2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(
4, path, edges, sentenceGraph) # remove for fast
#self.buildEdgeCombinations(path, edges, sentenceGraph, features) # remove for fast
#if edges != None:
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[0], edges[0][1]+edges[1][0], "t1", sentenceGraph) # remove for fast
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[-1], edges[len(path)-1][len(path)-2]+edges[len(path)-2][len(path)-1], "t2", sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(
path, edges, sentenceGraph)
self.multiEdgeFeatureBuilder.buildSentenceFeatures(
sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None)
if not "no_linear" in self.styles:
self.tokenFeatureBuilder.setFeatureVector(features)
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == token1:
token1Index = i
if sentenceGraph.tokens[i] == token2:
token2Index = i
linearPreTag = "linfw_"
if token1Index > token2Index:
token1Index, token2Index = token2Index, token1Index
linearPreTag = "linrv_"
self.tokenFeatureBuilder.buildLinearOrderFeatures(
token1Index, sentenceGraph, 2, 2, preTag="linTok1")
self.tokenFeatureBuilder.buildLinearOrderFeatures(
token2Index, sentenceGraph, 2, 2, preTag="linTok2")
# Before, middle, after
# self.tokenFeatureBuilder.buildTokenGrams(0, token1Index-1, sentenceGraph, "bf")
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, "bw")
# self.tokenFeatureBuilder.buildTokenGrams(token2Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, "af")
# before-middle, middle, middle-after
# self.tokenFeatureBuilder.buildTokenGrams(0, token2Index-1, sentenceGraph, linearPreTag+"bf", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, linearPreTag+"bw", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, linearPreTag+"af", max=2)
self.tokenFeatureBuilder.setFeatureVector(None)
if "random" in self.styles:
self.randomFeatureBuilder.setFeatureVector(features)
self.randomFeatureBuilder.buildRandomFeatures(100, 0.01)
self.randomFeatureBuilder.setFeatureVector(None)
if "genia_limits" in self.styles:
e1Type = entity1.get("type")
e2Type = entity2.get("type")
assert (entity1.get("isName") == "False")
if entity2.get("isName") == "True":
features[self.featureSet.getId(
"GENIA_target_protein")] = 1
else:
features[self.featureSet.getId(
"GENIA_nested_event")] = 1
if e1Type.find(
"egulation"
) != -1: # leave r out to avoid problems with capitalization
if entity2.get("isName") == "True":
features[self.featureSet.getId(
"GENIA_regulation_of_protein")] = 1
else:
features[self.featureSet.getId(
"GENIA_regulation_of_event")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
path = [token1, token2]
# define extra attributes
if int(path[0].attrib["id"].split("_")[-1]) < int(
path[-1].attrib["id"].split("_")[-1]):
#extra = {"xtype":"edge","type":"i","t1":path[0],"t2":path[-1]}
extra = {
"xtype": "ue",
"type": "i",
"t1": path[0].get("id"),
"t2": path[-1].get("id")
}
extra["deprev"] = False
else:
#extra = {"xtype":"edge","type":"i","t1":path[-1],"t2":path[0]}
extra = {
"xtype": "ue",
"type": "i",
"t1": path[-1].get("id"),
"t2": path[0].get("id")
}
extra["deprev"] = True
if entity1 != None:
extra["e1"] = entity1.get("id")
extra["l1"] = str(e1[1])
extra["d1"] = str(e1[2])[
0] # is a dummy node (an entity not in existing triggers)
if entity2 != None:
extra["e2"] = entity2.get("id")
extra["l2"] = str(e2[1])
extra["d2"] = str(e2[2])[
0] # is a dummy node (an entity not in existing triggers)
extra["categoryName"] = categoryName
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
if "binary" in self.styles:
if categoryName != "neg":
category = 1
else:
category = -1
categoryName = "i"
else:
category = self.classSet.getId(categoryName)
return (sentenceGraph.getSentenceId() + ".x" + str(exampleIndex),
category, features, extra)
class DirectEventExampleBuilder(ExampleBuilder):
def __init__(self, style=["typed","directed","headsOnly"], length=None, types=[], featureSet=None, classSet=None, gazetteer=None, pathGazetteer=None, negFrac=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert( classSet.getId("neg") == 1 )
if gazetteer != None:
print >> sys.stderr, "Loading gazetteer from", gazetteer
self.gazetteer=Gazetteer.loadGztr(gazetteer)
else:
print >> sys.stderr, "No gazetteer loaded"
self.gazetteer=None
self.pathGazetteer=None
self.pathGazetteerDependencies = None
self.pathGazetteerPairs = None
if pathGazetteer != None:
print >> sys.stderr, "Loading path gazetteer from", pathGazetteer
self.pathGazetteer=PathGazetteer.load(pathGazetteer)
self.pathGazetteerDependencies = PathGazetteer.getDependencies(self.pathGazetteer)
self.pathGazetteerPairs = PathGazetteer.getPairs(self.pathGazetteer)
else:
print >> sys.stderr, "No path gazetteer loaded"
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
self.styles = style
self.negFrac = negFrac
print >> sys.stderr, "Downsampling negatives to", negFrac
self.negRand = random.Random()
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if True:#"noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.triggerFeatureBuilder = TriggerFeatureBuilder(self.featureSet)
#self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
#if "ontology" in self.styles:
# self.multiEdgeFeatureBuilder.ontologyFeatureBuilder = BioInferOntologyFeatureBuilder(self.featureSet)
self.pathLengths = length
assert(self.pathLengths == None)
self.types = types
self.eventsByOrigId = {}
self.headTokensByOrigId = {}
self.interSentenceEvents = set()
self.examplesByEventOrigId = {}
self.skippedByType = {}
self.skippedByTypeAndReason = {}
self.builtByType = {}
self.gazMatchCache = {}
#self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls, input, output, parse, tokenization, style, idFileTag=None, gazetteer=None, pathGazetteer=None, negFrac=None):
classSet, featureSet = cls.getIdSets(idFileTag)
if style != None:
e = DirectEventExampleBuilder(style=style, classSet=classSet, featureSet=featureSet, gazetteer=gazetteer, pathGazetteer=pathGazetteer, negFrac=negFrac)
else:
e = DirectEventExampleBuilder(classSet=classSet, featureSet=featureSet, gazetteer=gazetteer, pathGazetteer=pathGazetteer, negFrac=negFrac)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
e.printStats()
def getGazetteerMatch(self, string):
if string in self.gazMatchCache:
return self.gazMatchCache[string]
origString = string
if "stem_gazetteer" in self.styles:
string = PorterStemmer.stem(string)
if string in self.gazetteer:
self.gazMatchCache[origString] = string
return string
elif string.find("-") != -1:
replaced = string.replace("-","")
else:
self.gazMatchCache[origString] = None
return None
if replaced in self.gazetteer:
self.gazMatchCache[origString] = replaced
return replaced
else:
splitted = string.rsplit("-",1)[-1]
if splitted in self.gazetteer:
self.gazMatchCache[origString] = splitted
return splitted
else:
self.gazMatchCache[origString] = None
return None
def isInGazetteer(self, string):
return self.getGazetteerMatch(string) != None
def printStats(self):
eventsByType = {}
for event in self.eventsByOrigId.values():
eventsByType[event.get("type")] = eventsByType.get(event.get("type"),0) + 1
f = open("missed-events", "wt")
missedEvents = {}
for key in self.examplesByEventOrigId.keys():
if self.examplesByEventOrigId[key] == 0:
if not missedEvents.has_key(self.eventsByOrigId[key].get("type")):
missedEvents[self.eventsByOrigId[key].get("type")] = []
missedEvents[self.eventsByOrigId[key].get("type")].append(key)
for key in sorted(missedEvents.keys()):
f.write(key + "\n")
for id in sorted(missedEvents[key]):
f.write(" " + id + " ")
if id in self.interSentenceEvents:
f.write("intersentence ")
text = self.headTokensByOrigId[id].get("text").lower()
if not self.isInGazetteer(text):
text = self.headTokensByOrigId[id].get("text").lower()
if "stem_gazetteer" in self.styles:
stemmed = PorterStemmer.stem(text)
f.write("not-in-gazetteer (" + text + " / " + stemmed +")" )
f.write("\n")
f.close()
print >> sys.stderr, "Example selection missed events (other, intersentence, non-gazetteer)"
for key in sorted(eventsByType.keys()):
inter = 0
other = 0
nongaz = 0
if missedEvents.has_key(key):
for id in missedEvents[key]:
tokText = self.headTokensByOrigId[id].get("text").lower()
if id in self.interSentenceEvents:
inter += 1
elif not self.isInGazetteer(tokText):
nongaz += 1
else:
other += 1
if inter == other == nongaz == 0:
print >> sys.stderr, " " + key + " (" + str(eventsByType[key]) + "): missed none"
else:
print >> sys.stderr, " " + key + " (" + str(eventsByType[key]) + "): " + str(other) + ", " + str(inter) + ", " + str(nongaz)
print >> sys.stderr, "Example generation (total, built/skipped)"
for key in sorted(list(set(self.skippedByType.keys() + self.builtByType.keys()))):
string = " " + key + ": (" + str(self.builtByType.get(key,0)+self.skippedByType.get(key,0)) + ", " + str(self.builtByType.get(key,0)) + "/" + str(self.skippedByType.get(key,0)) + ") ["
for key2 in sorted(self.skippedByTypeAndReason[key].keys()):
string += key2 + ":" + str(self.skippedByTypeAndReason[key][key2]) + " "
string += "]"
print >> sys.stderr, string
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
# def isPotentialGeniaInteraction(self, e1, e2):
# if e1.get("isName") == "True" and e2.get("isName") == "True":
# return False
# elif e1.get("isName") == "True" and e2.get("isName") == "False":
# return False
# else:
# return True
def getArgumentEntities(self, sentenceGraph, entityNode):
eId = entityNode.get("id")
assert(eId != None)
themeNodes = []
causeNodes = []
for edge in sentenceGraph.interactions:
if edge.get("e1") == eId:
edgeType = edge.get("type")
assert(edgeType in ["Theme", "Cause"]), edgeType
if edgeType == "Theme":
themeNodes.append( sentenceGraph.entitiesById[edge.get("e2")] )
elif edgeType == "Cause":
causeNodes.append( sentenceGraph.entitiesById[edge.get("e2")] )
return themeNodes, causeNodes
def makeGSEvents(self, sentenceGraph):
self.namedEntityHeadTokenIds = set()
self.gsEvents = {} # [token]->[event-type]->[1-n argument sets]
for token in sentenceGraph.tokens:
self.gsEvents[token] = {}
for entity in sentenceGraph.entities:
if entity.get("type") == "neg":
continue
elif entity.get("isName") == "True":
self.namedEntityHeadTokenIds.add(sentenceGraph.entityHeadTokenByEntity[entity].get("id"))
continue
eId = entity.get("id")
eOrigId = entity.get("origId")
assert not self.eventsByOrigId.has_key(eOrigId)
self.eventsByOrigId[eOrigId] = entity
if not self.examplesByEventOrigId.has_key(eOrigId):
self.examplesByEventOrigId[eOrigId] = 0
if len(sentenceGraph.interSentenceInteractions) > 0:
for interaction in sentenceGraph.interSentenceInteractions:
if interaction.get("e1") == eId:
self.interSentenceEvents.add(eOrigId)
eType = entity.get("type")
arguments = set()
for interaction in sentenceGraph.interactions:
if interaction.get("e1") == eId:
e2 = sentenceGraph.entitiesById[interaction.get("e2")]
e2TokenId = sentenceGraph.entityHeadTokenByEntity[e2].get("id")
arguments.add( (interaction.get("type"), e2TokenId ) )
#arguments.add( (interaction.get("type"), interaction.get("e2") ) )
arguments = tuple(sorted(list(arguments)))
eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
self.headTokensByOrigId[eOrigId] = eHeadToken
if not self.gsEvents[eHeadToken].has_key(eType):
self.gsEvents[eHeadToken][eType] = {}
if len(arguments) > 0:
if not self.gsEvents[eHeadToken][eType].has_key(arguments):
self.gsEvents[eHeadToken][eType][arguments] = []
self.gsEvents[eHeadToken][eType][arguments].append(eOrigId)
def getGSEventType(self, sentenceGraph, eHeadToken, themeTokens, causeTokens):
#eHeadToken = sentenceGraph.entityHeadTokenByEntity[entity]
#eType = entity.get("type")
if len(self.gsEvents[eHeadToken]) == 0:
return "neg", []
argumentSet = set()
for themeNode in themeTokens:
if themeNode != None:
argumentSet.add( ("Theme", themeNode.get("id")) )
for causeNode in causeTokens:
if causeNode != None:
argumentSet.add( ("Cause", causeNode.get("id")) )
argumentSet = tuple(sorted(list(argumentSet)))
gsTypes = set()
eventIds = []
for eventType in sorted(self.gsEvents[eHeadToken].keys()):
if argumentSet in self.gsEvents[eHeadToken][eventType].keys():
gsTypes.add(eventType)
eventIds.extend(self.gsEvents[eHeadToken][eventType][argumentSet])
if len(gsTypes) == 0:
return "neg", eventIds
elif len(gsTypes) == 1:
return list(gsTypes)[0], eventIds
else:
gsTypes = sorted(list(gsTypes))
string = gsTypes[0]
for gsType in gsTypes[1:]:
string += "---" + gsType
return string, eventIds
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def buildExamples(self, sentenceGraph):
self.makeGSEvents(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(resetCache=True)
self.triggerFeatureBuilder.initSentence(sentenceGraph)
examples = []
exampleIndex = 0
#undirected = sentenceGraph.dependencyGraph.to_undirected()
undirected = self.nxMultiDiGraphToUndirected(sentenceGraph.dependencyGraph)
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
eventTokens = []
nameTokens = []
gazCategories = {None:{"neg":-1}}
#stems = {}
for token in sentenceGraph.tokens:
gazText = self.getGazetteerMatch(token.get("text").lower())
if gazText != None:
gazCategories[token] = self.gazetteer[gazText]
else:
gazCategories[token] = {"neg":-1}
if token.get("id") in self.namedEntityHeadTokenIds:
nameTokens.append(token)
elif gazText != None:
eventTokens.append(token)
allTokens = eventTokens + nameTokens
#if len(nameTokens) == 0: # there can be no events in this sentence
# self.gsEvents = None
# return []
for token in eventTokens:
#gazCategories = self.gazetteer[token.get("text").lower()]
#print token.get("text").lower(), gazCategories
#multiargument = False
potentialRegulation = False
potentialBinding = False
for key in gazCategories[token].keys():
if key in ["Regulation","Positive_regulation","Negative_regulation"]:
#multiargument = True
potentialRegulation = True
break
for key in gazCategories[token].keys():
if key in ["Binding"]:
#multiargument = True
potentialBinding = True
break
if potentialRegulation:
combinations = combine.combine(allTokens, allTokens+[None])
else:
combinations = []
for t2 in nameTokens: #allTokens:
combinations.append( (t2, None) )
if potentialBinding:
for i in range(len(nameTokens) - 1):
for j in range(i+1, len(nameTokens)):
combinations.append( ((nameTokens[i],nameTokens[j]), None) )
for combination in combinations:
theme2Binding = False
if type(combination[0]) == types.ListType or type(combination[0]) == types.TupleType:
theme2Binding = True
categoryName, eventIds = self.getGSEventType(sentenceGraph, token, combination[0], [combination[1]])
else:
categoryName, eventIds = self.getGSEventType(sentenceGraph, token, [combination[0]], [combination[1]])
for id in eventIds:
self.examplesByEventOrigId[id] += 1
skip = False
s = self.skippedByTypeAndReason
if not s.has_key(categoryName):
s[categoryName] = {}
if gazCategories[token].get("neg",-1) > 0.99:
pass
if combination[0] == combination[1]:
pass #skip = True
if combination[0] == token or combination[1] == token:
if theme2Binding or gazCategories[combination[0]].get("Positive_regulation",-1) < 0:
skip = True
s[categoryName]["duparg"] = s[categoryName].get("duparg", 0) + 1
if combination[0] == None and combination[1] == None:
skip = True
s[categoryName]["noncmb"] = s[categoryName].get("noncmb", 0) + 1
validCat = self.isValidEvent(paths, sentenceGraph, token, combination)
if validCat != "OK": #not self.isValidEvent(paths, sentenceGraph, token, combination):
skip = True
#s[categoryName]["valid"] = s[categoryName].get("valid", 0) + 1
s[categoryName][validCat] = s[categoryName].get(validCat, 0) + 1
if len(nameTokens) == 0:
skip = True
s[categoryName]["non"] = s[categoryName].get("non", 0) + 1
if theme2Binding:
if gazCategories[combination[0][0]].get("neg",-1) > 0.99 or gazCategories[combination[0][1]].get("neg",-1) > 0.99:
skip = True
s[categoryName]["gazarg"] = s[categoryName].get("gazarg", 0) + 1
else:
if gazCategories[combination[0]].get("neg",-1) > 0.99 or gazCategories[combination[1]].get("neg",-1) > 0.99:
skip = True
s[categoryName]["gazarg"] = s[categoryName].get("gazarg", 0) + 1
if (skip and self.negFrac == None) or (skip and self.negFrac != None and categoryName == "neg"):
self.skippedByType[categoryName] = self.skippedByType.get(categoryName, 0) + 1
else:
if self.negFrac == None or categoryName != "neg" or (categoryName == "neg" and self.negRand.random() < self.negFrac):
self.builtByType[categoryName] = self.builtByType.get(categoryName, 0) + 1
if theme2Binding:
newExample = self.buildExample(exampleIndex, sentenceGraph, paths, token, combination[0], [combination[1]])
else:
newExample = self.buildExample(exampleIndex, sentenceGraph, paths, token, [combination[0]], [combination[1]])
if len(eventIds) > 0:
newExample[3]["numEv"] = str(len(eventIds))
examples.append( newExample )
exampleIndex += 1
self.gsEvents = None
return examples
def isValidEvent(self, paths, sentenceGraph, eventToken, argTokens):
# This one lets through Positive_regulations that are
# excluded from the duparg-rule
oneTokenEvent = True
for argToken in argTokens:
if argToken != None and eventToken != argToken:
oneTokenEvent = False
break
if oneTokenEvent:
return "OK" #True
if not paths.has_key(eventToken):
return "nopaths" #False
newArgTokens = []
for argToken in argTokens:
if type(argToken) == types.ListType or type(argToken) == types.TupleType:
newArgTokens.extend(argToken)
else:
newArgTokens.append(argToken)
argTokens = newArgTokens
oneArgValid = True
if False:
oneArgValid = False
for argToken in argTokens:
if argToken == None:
continue
if paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
else:
#print argToken, argToken.get("text")
#return False
continue
depPaths = self.multiEdgeFeatureBuilder.getEdgeCombinations(sentenceGraph.dependencyGraph, path)
validArg = False
for p in depPaths:
if p in self.pathGazetteer and self.pathGazetteer[p][0] > 0:
validArg = True
break
if validArg:
oneArgValid = True
# The first and last dependency of a path
if False:
oneEdgeValid = False
for argToken in argTokens:
if argToken == None:
continue
if paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
else:
#print argToken, argToken.get("text")
#return False
continue
depPaths = self.multiEdgeFeatureBuilder.getEdgeCombinations(sentenceGraph.dependencyGraph, path)
validArg = False
for p in depPaths:
p = p.replace("<","")
p = p.replace(">","")
p = p.split(".")
pair = (p[0], p[-1])
if pair in self.pathGazetteerPairs:
validArg = True
break
if validArg:
oneEdgeValid = True
break
if not oneEdgeValid:
return "pair"
# Event must not have unseen dependencies in any of its paths
if False:
for argToken in argTokens:
if argToken == None:
continue
if paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
else:
continue
deps = self.multiEdgeFeatureBuilder.getEdgeSet(sentenceGraph.dependencyGraph, path)
for d in deps:
if d[2].get("type") not in self.pathGazetteerDependencies:
#print "Unk", d[2].get("type")
return "unkdep"
# validArg = True
# for p in depPaths:
# if p in self.pathGazetteer and self.pathGazetteer[p][0] == 0:
# validArg = False
# break
# if not validArg:
# return False
if not oneArgValid:
return "novalidarg" #False
return "OK" #True
def setGazetteerFeatures(self, token, tag):
gazText = self.getGazetteerMatch(token.get("text").lower())
if gazText != None:
gazCategories = self.gazetteer[gazText]
for k,v in gazCategories.iteritems():
self.setFeature(tag+"gaz_event_value_"+k, v)
self.setFeature(tag+"gaz_event_"+k, 1)
if k.find("egulation") != -1:
self.setFeature(tag+"potReg", 1)
else:
self.setFeature(tag+"notInGaz", 1)
def buildExample(self, exampleIndex, sentenceGraph, paths, eventToken, themeTokens, causeTokens=None):
features = {}
self.features = features
categoryName, eventIds = self.getGSEventType(sentenceGraph, eventToken, themeTokens, causeTokens)
category = self.classSet.getId(categoryName)
potentialRegulation = False
eventTokenText = eventToken.get("text").lower()
gazText = self.getGazetteerMatch(eventTokenText)
gazCategories = self.gazetteer[gazText]
for k,v in gazCategories.iteritems():
if k.find("egulation") != -1:
potentialRegulation = True
self.setGazetteerFeatures(eventToken,"")
self.triggerFeatureBuilder.setFeatureVector(self.features)
self.triggerFeatureBuilder.tag = "trg_"
self.triggerFeatureBuilder.buildFeatures(eventToken)
themeEntities = []
hasTheme = False
if len(themeTokens) > 1:
self.setFeature("multiTheme", 1)
potentialRegulation = False
for themeToken in themeTokens:
if themeToken != None:
hasTheme = True
self.setGazetteerFeatures(themeToken,"theme_")
self.buildArgumentFeatures(sentenceGraph, paths, features, eventToken, themeToken, "theme_")
self.triggerFeatureBuilder.tag = "ttrg_"
self.triggerFeatureBuilder.buildFeatures(themeToken)
themeEntity = None
if sentenceGraph.entitiesByToken.has_key(themeToken):
for themeEntity in sentenceGraph.entitiesByToken[themeToken]:
if themeEntity.get("isName") == "True":
self.setFeature("themeProtein", 1)
if potentialRegulation:
self.setFeature("regulationThemeProtein", 1)
themeEntities.append(themeEntity)
break
if not features.has_key("themeProtein"):
self.setFeature("themeEvent", 1)
self.setFeature("nestingEvent", 1)
if potentialRegulation:
self.setFeature("regulationThemeEvent", 1)
if hasTheme:
self.setFeature("noTheme", 1)
causeEntities = []
hasCause = False
for causeToken in causeTokens:
if causeToken != None:
hasCause = True
self.setGazetteerFeatures(causeToken,"cause_")
self.buildArgumentFeatures(sentenceGraph, paths, features, eventToken, causeToken, "cause_")
self.triggerFeatureBuilder.tag = "ctrg_"
self.triggerFeatureBuilder.buildFeatures(causeToken)
causeEntity = None
if sentenceGraph.entitiesByToken.has_key(causeToken):
for causeEntity in sentenceGraph.entitiesByToken[causeToken]:
if causeEntity.get("isName") == "True":
self.setFeature("causeProtein", 1)
if potentialRegulation:
self.setFeature("regulationCauseProtein", 1)
causeEntities.append(causeEntity)
break
if not features.has_key("causeProtein"):
self.setFeature("causeEvent", 1)
self.setFeature("nestingEvent", 1)
if potentialRegulation:
self.setFeature("regulationCauseEvent", 1)
if not hasCause:
self.setFeature("noCause", 1)
self.triggerFeatureBuilder.tag = ""
self.triggerFeatureBuilder.setFeatureVector(None)
# Common features
# if e1Type.find("egulation") != -1: # leave r out to avoid problems with capitalization
# if entity2.get("isName") == "True":
# features[self.featureSet.getId("GENIA_regulation_of_protein")] = 1
# else:
# features[self.featureSet.getId("GENIA_regulation_of_event")] = 1
# define extra attributes
extra = {"xtype":"event","type":categoryName}
extra["et"] = eventToken.get("id")
if len(eventIds) > 0:
eventIds.sort()
extra["eids"] = ""
for eventId in eventIds:
extra["eids"] += str(eventId) + ","
extra["eids"] = extra["eids"][:-1]
for themeToken in themeTokens:
if themeToken != None:
if extra.has_key("tt"):
extra["tt"] = extra["tt"] + "," + themeToken.get("id")
else:
extra["tt"] = themeToken.get("id")
for themeEntity in themeEntities:
if extra.has_key("t"):
extra["t"] = extra["t"] + "," + themeEntity.get("id")
else:
extra["t"] = themeEntity.get("id")
for causeToken in causeTokens:
if causeToken != None:
extra["ct"] = causeTokens[0].get("id")
if len(causeEntities) > 0:
extra["c"] = causeEntities[0].get("id")
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
#assert (category == 1 or category == -1)
self.features = None
return (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra)
def buildArgumentFeatures(self, sentenceGraph, paths, features, eventToken, argToken, tag):
#eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
#argToken = sentenceGraph.entityHeadTokenByEntity[argNode]
self.multiEdgeFeatureBuilder.tag = tag
self.multiEdgeFeatureBuilder.setFeatureVector(features, None, None, False)
if eventToken != argToken and paths.has_key(eventToken) and paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
edges = self.multiEdgeFeatureBuilder.getEdges(sentenceGraph.dependencyGraph, path)
else:
path = [eventToken, argToken]
edges = None
# if not "disable_entity_features" in self.styles:
# # doesn't improve beyond 52.32
# self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
# # buildPathLengthFeatures 52.32 -> 51-51
# self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
# if not "disable_terminus_features" in self.styles:
# # didn't improve from 52.32
# self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
# 50.74 -> 52.32
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
# ngrams alone - 50.74
self.multiEdgeFeatureBuilder.buildPathGrams(2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(4, path, edges, sentenceGraph) # remove for fast
# disabling length 4 drops performance
# if not "disable_path_edge_features" in self.styles:
# self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(path, edges, sentenceGraph)
# self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
# buildSentenceFeatures seems to decrease performance by 8 %-points
self.multiEdgeFeatureBuilder.setFeatureVector(None, None, None, False)
self.multiEdgeFeatureBuilder.tag = ""
class AsymmetricEventExampleBuilder(ExampleBuilder):
def __init__(self, style=["typed","directed"], length=None, types=[], featureSet=None, classSet=None):
if featureSet == None:
featureSet = IdSet()
if classSet == None:
classSet = IdSet(1)
else:
classSet = classSet
assert( classSet.getId("neg") == 1 )
ExampleBuilder.__init__(self, classSet=classSet, featureSet=featureSet)
if style.find(",") != -1:
style = style.split(",")
self.styles = style
self.negFrac = None
self.posPairGaz = POSPairGazetteer()
for s in style:
if s.find("negFrac") != -1:
self.negFrac = float(s.split("_")[-1])
print >> sys.stderr, "Downsampling negatives to", self.negFrac
self.negRand = random.Random(15)
elif s.find("posPairGaz") != -1:
self.posPairGaz = POSPairGazetteer(loadFrom=s.split("_", 1)[-1])
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
self.triggerFeatureBuilder = TriggerFeatureBuilder(self.featureSet)
if "graph_kernel" in self.styles:
from FeatureBuilders.GraphKernelFeatureBuilder import GraphKernelFeatureBuilder
self.graphKernelFeatureBuilder = GraphKernelFeatureBuilder(self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.tokenFeatureBuilder = TokenFeatureBuilder(self.featureSet)
if "ontology" in self.styles:
self.multiEdgeFeatureBuilder.ontologyFeatureBuilder = BioInferOntologyFeatureBuilder(self.featureSet)
if "nodalida" in self.styles:
self.nodalidaFeatureBuilder = NodalidaFeatureBuilder(self.featureSet)
#IF LOCAL
if "bioinfer_limits" in self.styles:
self.bioinferOntologies = OntologyUtils.getBioInferTempOntology()
#self.bioinferOntologies = OntologyUtils.loadOntologies(OntologyUtils.g_bioInferFileName)
#ENDIF
self.pathLengths = length
assert(self.pathLengths == None)
self.types = types
if "random" in self.styles:
from FeatureBuilders.RandomFeatureBuilder import RandomFeatureBuilder
self.randomFeatureBuilder = RandomFeatureBuilder(self.featureSet)
#self.outFile = open("exampleTempFile.txt","wt")
@classmethod
def run(cls, input, output, parse, tokenization, style, idFileTag=None):
classSet, featureSet = cls.getIdSets(idFileTag)
if style != None:
e = cls(style=style, classSet=classSet, featureSet=featureSet)
else:
e = cls(classSet=classSet, featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
e.buildExamplesForSentences(sentences, output, idFileTag)
if "printClassIds" in e.styles:
print >> sys.stderr, e.classSet.Ids
def definePredictedValueRange(self, sentences, elementName):
self.multiEdgeFeatureBuilder.definePredictedValueRange(sentences, elementName)
def getPredictedValueRange(self):
return self.multiEdgeFeatureBuilder.predictedRange
def filterEdgesByType(self, edges, typesToInclude):
if len(typesToInclude) == 0:
return edges
edgesToKeep = []
for edge in edges:
if edge.get("type") in typesToInclude:
edgesToKeep.append(edge)
return edgesToKeep
def getCategoryNameFromTokens(self, sentenceGraph, t1, t2, directed=True):
types = set()
themeE1Types = set()
intEdges = []
if sentenceGraph.interactionGraph.has_edge(t1, t2):
intEdges = sentenceGraph.interactionGraph.get_edge_data(t1, t2, default={})
# NOTE: Only works if keys are ordered integers
for i in range(len(intEdges)):
types.add(intEdges[i]["element"].get("type"))
# if (not directed) and sentenceGraph.interactionGraph.has_edge(t2, t1):
# intEdgesReverse = sentenceGraph.interactionGraph.get_edge(t2, t1, default={})
# # NOTE: Only works if keys are ordered integers
# for i in range(len(intEdgesReverse)):
# intElement = intEdgesReverse[i]["element"]
# intType = intElement.get("type")
# types.add(intType)
# intEdges.extend(intEdgesReverse)
for i in range(len(intEdges)):
intElement = intEdges[i]["element"]
intType = intElement.get("type")
if intType == "Theme":
e1Entity = sentenceGraph.entitiesById[intElement.get("e1")]
themeE1Types.add(e1Entity.get("type"))
#types.add(intType)
if len(themeE1Types) != 0:
themeE1Types = list(themeE1Types)
themeE1Types.sort()
categoryName = ""
for name in themeE1Types:
if categoryName != "":
categoryName += "---"
categoryName += name
return categoryName
else:
types = list(types)
types.sort()
categoryName = ""
for name in types:
if categoryName != "":
categoryName += "---"
categoryName += name
if categoryName != "":
return categoryName
else:
return "neg"
def getCategoryName(self, sentenceGraph, e1, e2, directed=True):
interactions = sentenceGraph.getInteractions(e1, e2)
if not directed:
interactions.extend(sentenceGraph.getInteractions(e2, e1))
types = set()
for interaction in interactions:
types.add(interaction.attrib["type"])
types = list(types)
types.sort()
categoryName = ""
for name in types:
if categoryName != "":
categoryName += "---"
categoryName += name
if categoryName != "":
return categoryName
else:
return "neg"
def preProcessExamples(self, allExamples):
# Duplicates cannot be removed here, as they should only be removed from the training set. This is done
# in the classifier.
# if "no_duplicates" in self.styles:
# count = len(allExamples)
# print >> sys.stderr, " Removing duplicates,",
# allExamples = ExampleUtils.removeDuplicates(allExamples)
# print >> sys.stderr, "removed", count - len(allExamples)
if "normalize" in self.styles:
print >> sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def isPotentialGeniaInteraction(self, e1, e2):
if e1.get("isName") == "True":
return False
else:
return True
#IF LOCAL
def getBioInferParentType(self, eType):
if eType == "Physical_entity" or OntologyUtils.hasParent(eType, "Physical_entity", self.bioinferOntologies):
return "Physical"
elif eType == "Property_entity" or OntologyUtils.hasParent(eType, "Property_entity", self.bioinferOntologies):
return "Property"
elif OntologyUtils.hasParent(eType, "Relationship", self.bioinferOntologies):
return "Process"
else:
assert False, eType
# if self.bioinferOntologies["Entity"].has_key(eType):
# if OntologyUtils.hasParent(eType, "Physical_entity", self.bioinferOntologies):
# assert not OntologyUtils.hasParent(eType, "Property_entity", self.bioinferOntologies), eType
# return "Physical"
# else:
# assert OntologyUtils.hasParent(eType, "Property_entity", self.bioinferOntologies), eType
# return "Property"
#
# else:
# assert self.bioinferOntologies.has_key(eType), eType
# #assert OntologyUtils.hasParent(eType, "Process_entity", self.bioinferOntologies["Relationship"]), eType
# return "Process"
def isPotentialBioInferInteraction(self, e1, e2, categoryName):
e1Type = self.getBioInferParentType(e1.get("type"))
e2Type = self.getBioInferParentType(e2.get("type"))
if e1Type == "Process" or e1Type == "Property":
return True
elif e1Type == "Physical" and e2Type == "Physical":
return True
elif e1Type == "Physical" and e2Type == "Process": # hack
return True
else:
assert(categoryName == "neg"), categoryName + " category for " + e1Type + " and " + e2Type
return False
#ENDIF
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def buildExamples(self, sentenceGraph):
examples = []
exampleIndex = 0
clearGraph = sentenceGraph.getCleared()
#undirected = sentenceGraph.getUndirectedDependencyGraph()
undirected = self.nxMultiDiGraphToUndirected(sentenceGraph.dependencyGraph)
##undirected = sentenceGraph.dependencyGraph.to_undirected()
###undirected = NX10.MultiGraph(sentenceGraph.dependencyGraph) This didn't work
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
self.triggerFeatureBuilder.initSentence(clearGraph)
# Generate examples based on interactions between entities or interactions between tokens
if "entities" in self.styles:
loopRange = len(sentenceGraph.entities)
else:
loopRange = len(sentenceGraph.tokens)
#for i in range(loopRange-1):
for i in range(loopRange): # allow self-interactions
#for j in range(i+1,loopRange):
for j in range(i,loopRange): # allow self-interactions
eI = None
eJ = None
if "entities" in self.styles:
eI = sentenceGraph.entities[i]
eJ = sentenceGraph.entities[j]
tI = sentenceGraph.entityHeadTokenByEntity[eI]
tJ = sentenceGraph.entityHeadTokenByEntity[eJ]
#if "no_ne_interactions" in self.styles and eI.get("isName") == "True" and eJ.get("isName") == "True":
# continue
if eI.get("type") == "neg" or eJ.get("type") == "neg":
continue
else:
tI = sentenceGraph.tokens[i]
tJ = sentenceGraph.tokens[j]
# # only consider paths between entities (NOTE! entities, not only named entities)
# if "headsOnly" in self.styles:
# if (len(sentenceGraph.tokenIsEntityHead[tI]) == 0) or (len(sentenceGraph.tokenIsEntityHead[tJ]) == 0):
# continue
if "directed" in self.styles:
# define forward
if "entities" in self.styles:
categoryName = self.getCategoryName(sentenceGraph, eI, eJ, True)
else:
categoryName = self.getCategoryNameFromTokens(sentenceGraph, tI, tJ, True)
self.exampleStats.beginExample(categoryName)
if self.negFrac == None or categoryName != "neg" or (categoryName == "neg" and self.negRand.random() < self.negFrac):
makeExample = True
if ("genia_limits" in self.styles) and not self.isPotentialGeniaInteraction(eI, eJ):
makeExample = False
self.exampleStats.filter("genia_limits")
if self.posPairGaz.getNegFrac((tI.get("POS"), tJ.get("POS"))) == 1.0:
makeExample = False
self.exampleStats.filter("pos_pair")
if makeExample:
if not sentenceGraph.tokenIsName[tI]:
examples.append( self.buildExample(tI, tJ, paths, clearGraph, categoryName, exampleIndex, eI, eJ) )
exampleIndex += 1
else:
self.exampleStats.filter("genia_token_limits")
else:
self.exampleStats.filter("neg_frac")
self.exampleStats.endExample()
# define reverse
if "entities" in self.styles:
categoryName = self.getCategoryName(sentenceGraph, eJ, eI, True)
else:
categoryName = self.getCategoryNameFromTokens(sentenceGraph, tJ, tI, True)
self.exampleStats.beginExample(categoryName)
if self.negFrac == None or categoryName != "neg" or (categoryName == "neg" and self.negRand.random() < self.negFrac):
makeExample = True
if ("genia_limits" in self.styles) and not self.isPotentialGeniaInteraction(eJ, eI):
makeExample = False
self.exampleStats.filter("genia_limits")
if ("bioinfer_limits" in self.styles) and not self.isPotentialBioInferInteraction(eJ, eI, categoryName):
makeExample = False
self.exampleStats.filter("bioinfer_limits")
if self.posPairGaz.getNegFrac((tJ.get("POS"), tI.get("POS"))) == 1.0:
makeExample = False
self.exampleStats.filter("pos_pair")
if makeExample:
if not sentenceGraph.tokenIsName[tJ]:
examples.append( self.buildExample(tJ, tI, paths, clearGraph, categoryName, exampleIndex, eJ, eI) )
exampleIndex += 1
else:
self.exampleStats.filter("genia_token_limits")
else:
self.exampleStats.filter("neg_frac")
self.exampleStats.endExample()
# else:
# if "entities" in self.styles:
# categoryName = self.getCategoryName(sentenceGraph, eI, eJ, False)
# else:
# categoryName = self.getCategoryNameFromTokens(sentenceGraph, tI, tJ, False)
# forwardExample = self.buildExample(tI, tJ, paths, clearGraph, categoryName, exampleIndex, eI, eJ)
# if not "graph_kernel" in self.styles:
# reverseExample = self.buildExample(tJ, tI, paths, clearGraph, categoryName, exampleIndex, eJ, eI)
# forwardExample[2].update(reverseExample[2])
# examples.append(forwardExample)
# exampleIndex += 1
return examples
def buildExample(self, token1, token2, paths, sentenceGraph, categoryName, exampleIndex, entity1=None, entity2=None):
# define features
features = {}
if True: #token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
if token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
path = paths[token1][token2]
else:
path = [token1, token2]
assert(self.pathLengths == None)
if self.pathLengths == None or len(path)-1 in self.pathLengths:
if not "no_trigger":
self.triggerFeatureBuilder.setFeatureVector(self.features)
self.triggerFeatureBuilder.tag = "trg_t1_"
self.triggerFeatureBuilder.buildFeatures(eventToken)
self.triggerFeatureBuilder.tag = "trg_t2_"
self.triggerFeatureBuilder.buildFeatures(eventToken)
# if not "no_ontology" in self.styles:
# self.ontologyFeatureBuilder.setFeatureVector(features)
# self.ontologyFeatureBuilder.buildOntologyFeaturesForPath(sentenceGraph, path)
# self.ontologyFeatureBuilder.setFeatureVector(None)
if "graph_kernel" in self.styles or not "no_dependency" in self.styles:
if token1 != token2 and paths.has_key(token1) and paths[token1].has_key(token2):
edges = self.multiEdgeFeatureBuilder.getEdges(sentenceGraph.dependencyGraph, path)
else:
edges = None
if "graph_kernel" in self.styles:
self.graphKernelFeatureBuilder.setFeatureVector(features, entity1, entity2)
self.graphKernelFeatureBuilder.buildGraphKernelFeatures(sentenceGraph, path, edges)
self.graphKernelFeatureBuilder.setFeatureVector(None)
if "entity_type" in self.styles:
features[self.featureSet.getId("e1_"+entity1.attrib["type"])] = 1
features[self.featureSet.getId("e2_"+entity2.attrib["type"])] = 1
features[self.featureSet.getId("distance_"+str(len(path)))] = 1
if not "no_dependency" in self.styles:
if token1 == token2:
features[self.featureSet.getId("tokenSelfLoop")] = 1
self.multiEdgeFeatureBuilder.setFeatureVector(features, entity1, entity2)
#self.multiEdgeFeatureBuilder.buildStructureFeatures(sentenceGraph, paths) # remove for fast
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
if not "disable_terminus_features" in self.styles:
self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(4, path, edges, sentenceGraph) # remove for fast
#self.buildEdgeCombinations(path, edges, sentenceGraph, features) # remove for fast
#if edges != None:
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[0], edges[0][1]+edges[1][0], "t1", sentenceGraph) # remove for fast
# self.multiEdgeFeatureBuilder.buildTerminusFeatures(path[-1], edges[len(path)-1][len(path)-2]+edges[len(path)-2][len(path)-1], "t2", sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(path, edges, sentenceGraph)
self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None)
if "nodalida" in self.styles:
self.nodalidaFeatureBuilder.setFeatureVector(features, entity1, entity2)
shortestPaths = self.nodalidaFeatureBuilder.buildShortestPaths(sentenceGraph.dependencyGraph, path)
print shortestPaths
if len(shortestPaths) > 0:
self.nodalidaFeatureBuilder.buildNGrams(shortestPaths, sentenceGraph)
self.nodalidaFeatureBuilder.setFeatureVector(None)
if not "no_linear" in self.styles:
self.tokenFeatureBuilder.setFeatureVector(features)
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == token1:
token1Index = i
if sentenceGraph.tokens[i] == token2:
token2Index = i
linearPreTag = "linfw_"
if token1Index > token2Index:
token1Index, token2Index = token2Index, token1Index
linearPreTag = "linrv_"
self.tokenFeatureBuilder.buildLinearOrderFeatures(token1Index, sentenceGraph, 2, 2, preTag="linTok1")
self.tokenFeatureBuilder.buildLinearOrderFeatures(token2Index, sentenceGraph, 2, 2, preTag="linTok2")
# Before, middle, after
# self.tokenFeatureBuilder.buildTokenGrams(0, token1Index-1, sentenceGraph, "bf")
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, "bw")
# self.tokenFeatureBuilder.buildTokenGrams(token2Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, "af")
# before-middle, middle, middle-after
# self.tokenFeatureBuilder.buildTokenGrams(0, token2Index-1, sentenceGraph, linearPreTag+"bf", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, token2Index-1, sentenceGraph, linearPreTag+"bw", max=2)
# self.tokenFeatureBuilder.buildTokenGrams(token1Index+1, len(sentenceGraph.tokens)-1, sentenceGraph, linearPreTag+"af", max=2)
self.tokenFeatureBuilder.setFeatureVector(None)
if "random" in self.styles:
self.randomFeatureBuilder.setFeatureVector(features)
self.randomFeatureBuilder.buildRandomFeatures(100, 0.01)
self.randomFeatureBuilder.setFeatureVector(None)
if "genia_limits" in self.styles:
e1Type = entity1.get("type")
e2Type = entity2.get("type")
assert(entity1.get("isName") == "False")
if entity2.get("isName") == "True":
features[self.featureSet.getId("GENIA_target_protein")] = 1
else:
features[self.featureSet.getId("GENIA_nested_event")] = 1
if e1Type.find("egulation") != -1: # leave r out to avoid problems with capitalization
if entity2.get("isName") == "True":
features[self.featureSet.getId("GENIA_regulation_of_protein")] = 1
else:
features[self.featureSet.getId("GENIA_regulation_of_event")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
else:
features[self.featureSet.getId("always_negative")] = 1
if "subset" in self.styles:
features[self.featureSet.getId("out_of_scope")] = 1
path = [token1, token2]
self.triggerFeatureBuilder.tag = ""
self.triggerFeatureBuilder.setFeatureVector(None)
# define extra attributes
# if int(path[0].attrib["id"].split("_")[-1]) < int(path[-1].attrib["id"].split("_")[-1]):
# #extra = {"xtype":"edge","type":"i","t1":path[0],"t2":path[-1]}
# extra = {"xtype":"asym","type":"i","t1":path[0].get("id"),"t2":path[-1].get("id")}
# extra["deprev"] = False
# else:
# #extra = {"xtype":"edge","type":"i","t1":path[-1],"t2":path[0]}
# extra = {"xtype":"asym","type":"i","t1":path[-1].get("id"),"t2":path[0].get("id")}
# extra["deprev"] = True
extra = {"xtype":"asym","type":"i","t1":token1.get("id"),"t2":token2.get("id")}
if entity1 != None:
#extra["e1"] = entity1
extra["e1"] = entity1.get("id")
if entity2 != None:
#extra["e2"] = entity2
extra["e2"] = entity2.get("id")
extra["categoryName"] = categoryName
sentenceOrigId = sentenceGraph.sentenceElement.get("origId")
if sentenceOrigId != None:
extra["SOID"] = sentenceOrigId
# make example
if "binary" in self.styles:
if categoryName != "neg":
category = 1
else:
category = -1
categoryName = "i"
else:
category = self.classSet.getId(categoryName)
return (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra)
class Round2TriggerExampleBuilder(ExampleBuilder):
def nxMultiDiGraphToUndirected(self, graph):
undirected = NX10.MultiGraph(name=graph.name)
undirected.add_nodes_from(graph)
undirected.add_edges_from(graph.edges_iter())
return undirected
def getPredictionStrength(self, element):
eType = element.get("type")
predictions = element.get("predictions")
if predictions == None:
return 0
predictions = predictions.split(",")
for prediction in predictions:
predClass, predStrength = prediction.split(":")
if predClass == eType:
predStrength = float(predStrength)
return predStrength
return 0
def getInteractionEdgeLengths(self, sentenceGraph, paths):
"""
Return dependency and linear length of all interaction edges
(measured between the two tokens).
"""
interactionLengths = {}
for interaction in sentenceGraph.interactions:
# Calculated interaction edge dep and lin length
e1 = sentenceGraph.entitiesById[interaction.get("e1")]
e2 = sentenceGraph.entitiesById[interaction.get("e2")]
t1 = sentenceGraph.entityHeadTokenByEntity[e1]
t2 = sentenceGraph.entityHeadTokenByEntity[e2]
# Get dep path length
if t1 != t2 and paths.has_key(t1) and paths[t1].has_key(t2):
pathLength = len(paths[t1][t2])
else: # no dependencyPath
pathLength = 999999 # more than any real path
# Linear distance
t1Pos = -1
t2Pos = -1
for i in range(len(sentenceGraph.tokens)):
if sentenceGraph.tokens[i] == t1:
t1Pos = i
if t2Pos != -1:
break
if sentenceGraph.tokens[i] == t2:
t2Pos = i
if t1Pos != -1:
break
linLength = abs(t1Pos - t2Pos)
interactionLengths[interaction] = (interaction, pathLength, linLength, t2Pos)
return interactionLengths
def __init__(self, style=None, classSet=None, featureSet=None, gazetteerFileName=None, skiplist=None):
if classSet == None:
classSet = IdSet(1)
assert classSet.getId("neg") == 1
if featureSet == None:
featureSet = IdSet()
ExampleBuilder.__init__(self, classSet, featureSet)
# gazetteerFileName="/usr/share/biotext/GeniaChallenge/SharedTaskTriggerTest/gazetteer-train"
if gazetteerFileName != None:
self.gazetteer = Gazetteer.loadGztr(gazetteerFileName)
print >>sys.stderr, "Loaded gazetteer from", gazetteerFileName
else:
print >>sys.stderr, "No gazetteer loaded"
self.gazetteer = None
self.styles = style
self.skiplist = set()
if skiplist != None:
f = open(skiplist, "rt")
for line in f.readlines():
self.skiplist.add(line.strip())
f.close()
self.styles = [
"trigger_features",
"typed",
"directed",
"no_linear",
"entities",
"genia_limits",
"noMasking",
"maxFeatures",
]
self.multiEdgeFeatureBuilder = MultiEdgeFeatureBuilder(self.featureSet)
if "graph_kernel" in self.styles:
from FeatureBuilders.GraphKernelFeatureBuilder import GraphKernelFeatureBuilder
self.graphKernelFeatureBuilder = GraphKernelFeatureBuilder(self.featureSet)
if "noAnnType" in self.styles:
self.multiEdgeFeatureBuilder.noAnnType = True
if "noMasking" in self.styles:
self.multiEdgeFeatureBuilder.maskNamedEntities = False
if "maxFeatures" in self.styles:
self.multiEdgeFeatureBuilder.maximum = True
self.triggerFeatureBuilder = TriggerFeatureBuilder(self.featureSet)
@classmethod
def run(cls, input, gold, output, parse, tokenization, style, idFileTag=None, append=False):
"""
An interface for running the example builder without needing to create a class
"""
classSet, featureSet = cls.getIdSets(idFileTag)
if style != None:
e = Round2TriggerExampleBuilder(style=style, classSet=classSet, featureSet=featureSet)
else:
e = Round2TriggerExampleBuilder(classSet=classSet, featureSet=featureSet)
sentences = cls.getSentences(input, parse, tokenization)
if gold != None:
goldSentences = cls.getSentences(gold, parse, tokenization)
else:
goldSentences = None
e.buildExamplesForSentences(sentences, goldSentences, output, idFileTag, append=append)
def buildExamplesForSentences(self, sentences, goldSentences, output, idFileTag=None, append=False):
examples = []
counter = ProgressCounter(len(sentences), "Build examples")
if append:
outfile = open(output, "at")
else:
outfile = open(output, "wt")
exampleCount = 0
for i in range(len(sentences)):
sentence = sentences[i]
goldSentence = [None]
if goldSentences != None:
goldSentence = goldSentences[i]
counter.update(1, "Building examples (" + sentence[0].getSentenceId() + "): ")
examples = self.buildExamples(sentence[0], goldSentence[0], append=append)
exampleCount += len(examples)
examples = self.preProcessExamples(examples)
ExampleUtils.appendExamples(examples, outfile)
outfile.close()
print >>sys.stderr, "Examples built:", exampleCount
print >>sys.stderr, "Features:", len(self.featureSet.getNames())
# IF LOCAL
if self.exampleStats.getExampleCount() > 0:
self.exampleStats.printStats()
# ENDIF
# Save Ids
if idFileTag != None:
print >>sys.stderr, "Saving class names to", idFileTag + ".class_names"
self.classSet.write(idFileTag + ".class_names")
print >>sys.stderr, "Saving feature names to", idFileTag + ".feature_names"
self.featureSet.write(idFileTag + ".feature_names")
def preProcessExamples(self, allExamples):
if "normalize" in self.styles:
print >>sys.stderr, " Normalizing feature vectors"
ExampleUtils.normalizeFeatureVectors(allExamples)
return allExamples
def getMergedEntityType(self, entities):
"""
If a single token belongs to multiple entities of different types,
a new, composite type is defined. This type is the alphabetically
ordered types of these entities joined with '---'.
"""
types = set()
for entity in entities:
types.add(entity.get("type"))
types = list(types)
types.sort()
typeString = ""
for type in types:
if type == "Protein" and "all_tokens" in self.styles:
continue
if typeString != "":
typeString += "---"
typeString += type
if typeString == "":
return "neg"
if "limit_merged_types" in self.styles:
if typeString.find("---") != -1:
if typeString == "Gene_expression---Positive_regulation":
return typeString
else:
return typeString.split("---")[0]
else:
return typeString
return typeString
def getTokenFeatures(self, token, sentenceGraph):
"""
Returns a list of features based on the attributes of a token.
These can be used to define more complex features.
"""
# These features are cached when this method is first called
# for a token.
if self.tokenFeatures.has_key(token):
return self.tokenFeatures[token]
tokTxt = sentenceGraph.getTokenText(token)
features = {}
features["_txt_" + tokTxt] = 1
# F 69.35 -> 68.22
# normalizedText = tokTxt.replace("-","").replace("/","").replace(",","").replace("\\","").replace(" ","").lower()
# features["_norTxt_"+normalizedText]=1
# features["_norStem_" + PorterStemmer.stem(normalizedText)]=1
features["_POS_" + token.get("POS")] = 1
if sentenceGraph.tokenIsName[token]:
features["_isName"] = 1
for entity in sentenceGraph.tokenIsEntityHead[token]:
if entity.get("isName") == "True":
features["_annType_" + entity.get("type")] = 1
# Filip's gazetteer based features (can be used separately from exclude_gazetteer)
if "gazetteer_features" in self.styles:
tokTxtLower = tokTxt.lower()
if "stem_gazetteer" in self.styles:
tokTxtLower = PorterStemmer.stem(tokTxtLower)
if self.gazetteer and tokTxtLower in self.gazetteer:
for label, weight in self.gazetteer[tokTxtLower].items():
features["_knownLabel_" + label] = weight # 1 performs slightly worse
self.tokenFeatures[token] = features
return features
def buildLinearOrderFeatures(self, sentenceGraph, index, tag, features):
"""
Linear features are built by marking token features with a tag
that defines their relative position in the linear order.
"""
tag = "linear_" + tag
for tokenFeature, w in self.getTokenFeatures(sentenceGraph.tokens[index], sentenceGraph).iteritems():
features[self.featureSet.getId(tag + tokenFeature)] = w
def buildExamples(self, sentenceGraph, goldGraph, append=False):
examples = self.buildExamplesInner(sentenceGraph, goldGraph)
entityCounts = {}
exampleCounts = {}
for entity in sentenceGraph.entities:
eType = entity.get("type")
if eType == "Protein":
continue
if not entityCounts.has_key(eType):
entityCounts[eType] = 0
exampleCounts[eType] = 0
entityCounts[eType] += 1
for example in examples:
eTypes = self.classSet.getName(example[1]).split("---")
for eType in eTypes:
if not exampleCounts.has_key(eType):
exampleCounts[eType] = 0
exampleCounts[eType] += 1
# for key in sorted(entityCounts.keys()):
# if entityCounts[key] != exampleCounts[key]:
# print >> sys.stderr, "Warning, sentence", sentenceGraph.getSentenceId(), "example", key, "diff", entityCounts[key] - exampleCounts[key]
return examples
def buildExamplesInner(self, sentenceGraph, goldGraph):
"""
Build one example for each token of the sentence
"""
if sentenceGraph.sentenceElement.get("origId") in self.skiplist:
print >>sys.stderr, "Skipping sentence", sentenceGraph.sentenceElement.get("origId")
return []
self.multiEdgeFeatureBuilder.setFeatureVector(resetCache=True)
self.triggerFeatureBuilder.initSentence(sentenceGraph)
undirected = self.nxMultiDiGraphToUndirected(sentenceGraph.dependencyGraph)
paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
# Get argument order
self.interactionLengths = self.getInteractionEdgeLengths(sentenceGraph, paths)
self.interactionLengths = self.interactionLengths.values()
self.interactionLengths.sort(compareInteractionPrecedence)
# Map tokens to entities
tokenByOffset = {}
for i in range(len(sentenceGraph.tokens)):
token = sentenceGraph.tokens[i]
if goldGraph != None:
goldToken = goldGraph.tokens[i]
assert token.get("id") == goldToken.get("id") and token.get("charOffset") == goldToken.get("charOffset")
tokenByOffset[token.get("charOffset")] = token.get("id")
# Map gold entities to their head offsets
goldEntitiesByOffset = {}
for token in sentenceGraph.tokens:
goldEntitiesByOffset[token.get("charOffset")] = []
entityToGold = {}
for entity in sentenceGraph.entities:
entityToGold[entity] = []
if goldGraph != None:
for entity in goldGraph.entities:
offset = entity.get("headOffset")
assert offset != None
goldEntitiesByOffset[offset].append(entity)
# Map predicted entities to gold entities
for entity in sentenceGraph.entities:
eType = entity.get("type")
eOffset = entity.get("headOffset")
for goldEntity in goldEntitiesByOffset[eOffset]:
if goldEntity.get("type") == eType:
entityToGold[entity].append(goldEntity)
# Map entities to interactions
# interactionsByEntityId = {}
# for entity in sentenceGraph.entities:
# interactionsByEntityId[entity.get("id")] = []
# Map tokens to interactions
interactionsByToken = {}
for token in sentenceGraph.tokens:
interactionsByToken[token] = []
for interactionTuple in self.interactionLengths:
interaction = interactionTuple[0]
if interaction.get("type") == "neg":
continue
e1Id = interaction.get("e1")
token = sentenceGraph.entityHeadTokenByEntity[sentenceGraph.entitiesById[e1Id]]
interactionsByToken[token].append(interaction)
examples = []
exampleIndex = 0
self.tokenFeatures = {}
# namedEntityNorStrings = set()
namedEntityHeadTokens = []
if not "names" in self.styles:
namedEntityCount = 0
for entity in sentenceGraph.entities:
if entity.get("isName") == "True": # known data which can be used for features
namedEntityCount += 1
# namedEntityNorStrings.add( entity.get("text").replace("-","").replace("/","").replace(",","").replace("\\","").replace(" ","").lower() )
namedEntityCountFeature = "nameCount_" + str(namedEntityCount)
# if namedEntityCount == 0: # no names, no need for triggers
# return []
if "pos_pairs" in self.styles:
namedEntityHeadTokens = self.getNamedEntityHeadTokens(sentenceGraph)
# neFeatures = {} # F: 69.35 -> 69.14
# for norString in namedEntityNorStrings:
# neFeatures[self.featureSet.getId("norNE_" + norString)] = 1
bagOfWords = {}
for token in sentenceGraph.tokens:
text = "bow_" + token.get("text")
if not bagOfWords.has_key(text):
bagOfWords[text] = 0
bagOfWords[text] += 1
if sentenceGraph.tokenIsName[token]:
text = "ne_" + text
if not bagOfWords.has_key(text):
bagOfWords[text] = 0
bagOfWords[text] += 1
bowFeatures = {}
for k, v in bagOfWords.iteritems():
bowFeatures[self.featureSet.getId(k)] = v
self.inEdgesByToken = {}
self.outEdgesByToken = {}
self.edgeSetByToken = {}
for token in sentenceGraph.tokens:
inEdges = sentenceGraph.dependencyGraph.in_edges(token, data=True)
fixedInEdges = []
for edge in inEdges:
fixedInEdges.append((edge[0], edge[1], edge[2]["element"]))
inEdges = fixedInEdges
inEdges.sort(compareDependencyEdgesById)
self.inEdgesByToken[token] = inEdges
outEdges = sentenceGraph.dependencyGraph.out_edges(token, data=True)
fixedOutEdges = []
for edge in outEdges:
fixedOutEdges.append((edge[0], edge[1], edge[2]["element"]))
outEdges = fixedOutEdges
outEdges.sort(compareDependencyEdgesById)
self.outEdgesByToken[token] = outEdges
self.edgeSetByToken[token] = set(inEdges + outEdges)
for i in range(len(sentenceGraph.tokens)):
token = sentenceGraph.tokens[i]
# Recognize only non-named entities (i.e. interaction words)
if sentenceGraph.tokenIsName[token] and not "names" in self.styles and not "all_tokens" in self.styles:
continue
# CLASS
# if len(sentenceGraph.tokenIsEntityHead[token]) > 0:
# category = self.classSet.getId(self.getMergedEntityType(sentenceGraph.tokenIsEntityHead[token]))
# else:
# category = 1
offset = token.get("charOffset")
if len(goldEntitiesByOffset[offset]) > 0:
category = self.classSet.getId(self.getMergedEntityType(goldEntitiesByOffset[offset]))
else:
category = 1
tokenText = token.get("text").lower()
if "stem_gazetteer" in self.styles:
tokenText = PorterStemmer.stem(tokenText)
if ("exclude_gazetteer" in self.styles) and self.gazetteer and tokenText not in self.gazetteer:
features = {}
features[self.featureSet.getId("exclude_gazetteer")] = 1
extra = {"xtype": "token", "t": token.get("id"), "excluded": "True"}
examples.append((sentenceGraph.getSentenceId() + ".x" + str(exampleIndex), category, features, extra))
exampleIndex += 1
continue
# FEATURES
features = {}
self.features = features
if not "names" in self.styles:
features[self.featureSet.getId(namedEntityCountFeature)] = 1
# for k,v in bagOfWords.iteritems():
# features[self.featureSet.getId(k)] = v
# pre-calculate bow _features_
features.update(bowFeatures)
# features.update(neFeatures)
# for j in range(len(sentenceGraph.tokens)):
# text = "bow_" + sentenceGraph.tokens[j].get("text")
# if j < i:
# features[self.featureSet.getId("bf_" + text)] = 1
# elif j > i:
# features[self.featureSet.getId("af_" + text)] = 1
# Main features
text = token.get("text")
features[self.featureSet.getId("txt_" + text)] = 1
features[self.featureSet.getId("POS_" + token.get("POS"))] = 1
stem = PorterStemmer.stem(text)
features[self.featureSet.getId("stem_" + stem)] = 1
features[self.featureSet.getId("nonstem_" + text[len(stem) :])] = 1
# Normalized versions of the string (if same as non-normalized, overlap without effect)
normalizedText = (
text.replace("-", "").replace("/", "").replace(",", "").replace("\\", "").replace(" ", "").lower()
)
if normalizedText == "bound": # should be for all irregular verbs
normalizedText = "bind"
features[self.featureSet.getId("txt_" + normalizedText)] = 1
norStem = PorterStemmer.stem(normalizedText)
features[self.featureSet.getId("stem_" + norStem)] = 1
features[self.featureSet.getId("nonstem_" + normalizedText[len(norStem) :])] = 1
if "gazetteer_features_maintoken" in self.styles:
tokTxtLower = text.lower()
if "stem_gazetteer" in self.styles:
tokTxtLower = PorterStemmer.stem(tokTxtLower)
if self.gazetteer and tokTxtLower in self.gazetteer:
for label, weight in self.gazetteer[tokTxtLower].items():
features[self.featureSet.getId("gaz_knownLabel_" + label)] = weight # 1 performs slightly worse
# Linear order features
# for index in [-3,-2,-1,1,2,3,4,5]: # 69.35 -> 68.97
for index in [-3, -2, -1, 1, 2, 3]:
if i + index > 0 and i + index < len(sentenceGraph.tokens):
self.buildLinearOrderFeatures(sentenceGraph, i + index, str(index), features)
# Content
if i > 0 and text[0].isalpha() and text[0].isupper():
features[self.featureSet.getId("upper_case_start")] = 1
for j in range(len(text)):
if j > 0 and text[j].isalpha() and text[j].isupper():
features[self.featureSet.getId("upper_case_middle")] = 1
# numbers and special characters
if text[j].isdigit():
features[self.featureSet.getId("has_digits")] = 1
if j > 0 and text[j - 1] == "-":
features[self.featureSet.getId("has_hyphenated_digit")] = 1
elif text[j] == "-":
features[self.featureSet.getId("has_hyphen")] = 1
elif text[j] == "/":
features[self.featureSet.getId("has_fslash")] = 1
elif text[j] == "\\":
features[self.featureSet.getId("has_bslash")] = 1
# duplets
if j > 0:
features[self.featureSet.getId("dt_" + text[j - 1 : j + 1].lower())] = 1
# triplets
if j > 1:
features[self.featureSet.getId("tt_" + text[j - 2 : j + 1].lower())] = 1
# Attached edges (Hanging in and out edges)
t1InEdges = self.inEdgesByToken[token]
for edge in t1InEdges:
edgeType = edge[2].get("type")
features[self.featureSet.getId("t1HIn_" + edgeType)] = 1
features[self.featureSet.getId("t1HIn_" + edge[0].get("POS"))] = 1
features[self.featureSet.getId("t1HIn_" + edgeType + "_" + edge[0].get("POS"))] = 1
tokenText = sentenceGraph.getTokenText(edge[0])
features[self.featureSet.getId("t1HIn_" + tokenText)] = 1
features[self.featureSet.getId("t1HIn_" + edgeType + "_" + tokenText)] = 1
t1OutEdges = self.outEdgesByToken[token]
for edge in t1OutEdges:
edgeType = edge[2].get("type")
features[self.featureSet.getId("t1HOut_" + edgeType)] = 1
features[self.featureSet.getId("t1HOut_" + edge[1].get("POS"))] = 1
features[self.featureSet.getId("t1HOut_" + edgeType + "_" + edge[1].get("POS"))] = 1
tokenText = sentenceGraph.getTokenText(edge[1])
features[self.featureSet.getId("t1HOut_" + tokenText)] = 1
features[self.featureSet.getId("t1HOut_" + edgeType + "_" + tokenText)] = 1
extra = {"xtype": "token", "t": token.get("id")}
examples.append((sentenceGraph.getSentenceId() + ".x" + str(exampleIndex), category, features, extra))
exampleIndex += 1
# chains
self.buildChains(token, sentenceGraph, features)
if "pos_pairs" in self.styles:
self.buildPOSPairs(token, namedEntityHeadTokens, features)
self.buildPredictionFeatures(sentenceGraph, paths, token, interactionsByToken[token])
return examples
def buildChains(self, token, sentenceGraph, features, depthLeft=3, chain="", visited=None):
if depthLeft == 0:
return
strDepthLeft = "dist_" + str(depthLeft)
if visited == None:
visited = set()
inEdges = self.inEdgesByToken[token]
outEdges = self.outEdgesByToken[token]
edgeSet = visited.union(self.edgeSetByToken[token])
for edge in inEdges:
if not edge in visited:
edgeType = edge[2].get("type")
features[self.featureSet.getId("dep_" + strDepthLeft + edgeType)] = 1
nextToken = edge[0]
for tokenFeature, w in self.getTokenFeatures(nextToken, sentenceGraph).iteritems():
features[self.featureSet.getId(strDepthLeft + tokenFeature)] = w
# for entity in sentenceGraph.tokenIsEntityHead[nextToken]:
# if entity.get("isName") == "True":
# features[self.featureSet.getId("name_dist_"+strDepthLeft)] = 1
# features[self.featureSet.getId("name_dist_"+strDepthLeft+entity.get("type"))] = 1
# features[self.featureSet.getId("POS_dist_"+strDepthLeft+nextToken.get("POS"))] = 1
# tokenText = sentenceGraph.getTokenText(nextToken)
# features[self.featureSet.getId("text_dist_"+strDepthLeft+tokenText)] = 1
if sentenceGraph.tokenIsName[nextToken]:
features[self.featureSet.getId("name_chain_dist_" + strDepthLeft + chain + "-frw_" + edgeType)] = 1
features[self.featureSet.getId("chain_dist_" + strDepthLeft + chain + "-frw_" + edgeType)] = 1
self.buildChains(nextToken, sentenceGraph, features, depthLeft - 1, chain + "-frw_" + edgeType, edgeSet)
for edge in outEdges:
if not edge in visited:
edgeType = edge[2].get("type")
features[self.featureSet.getId("dep_dist_" + strDepthLeft + edgeType)] = 1
nextToken = edge[1]
for tokenFeature, w in self.getTokenFeatures(nextToken, sentenceGraph).iteritems():
features[self.featureSet.getId(strDepthLeft + tokenFeature)] = w
# for entity in sentenceGraph.tokenIsEntityHead[nextToken]:
# if entity.get("isName") == "True":
# features[self.featureSet.getId("name_dist_"+strDepthLeft)] = 1
# features[self.featureSet.getId("name_dist_"+strDepthLeft+entity.get("type"))] = 1
# features[self.featureSet.getId("POS_dist_"+strDepthLeft+nextToken.get("POS"))] = 1
# tokenText = sentenceGraph.getTokenText(nextToken)
# features[self.featureSet.getId("text_dist_"+strDepthLeft+tokenText)] = 1
if sentenceGraph.tokenIsName[nextToken]:
features[self.featureSet.getId("name_chain_dist_" + strDepthLeft + chain + "-rev_" + edgeType)] = 1
features[self.featureSet.getId("chain_dist_" + strDepthLeft + chain + "-rev_" + edgeType)] = 1
self.buildChains(nextToken, sentenceGraph, features, depthLeft - 1, chain + "-rev_" + edgeType, edgeSet)
def getNamedEntityHeadTokens(self, sentenceGraph):
headTokens = []
for entity in sentenceGraph.entities:
if entity.get("isName") == "True": # known data which can be used for features
headTokens.append(sentenceGraph.entityHeadTokenByEntity[entity])
return headTokens
def buildPOSPairs(self, token, namedEntityHeadTokens, features):
tokenPOS = token.get("POS")
assert tokenPOS != None
for headToken in namedEntityHeadTokens:
headPOS = headToken.get("POS")
features[self.featureSet.getId("POS_pair_NE_" + tokenPOS + "-" + headPOS)] = 1
######################################################
# Unmerging-style features
######################################################
def buildPredictionFeatures(self, sentenceGraph, paths, token, interactions): # themeEntities, causeEntities=None):
# NOTE!!!! TODO
# add also features for arguments present, but not in this combination
self.buildInterArgumentBagOfWords(interactions, sentenceGraph)
if sentenceGraph.entitiesByToken.has_key(token):
for eventEntity in sentenceGraph.entitiesByToken[token]:
eventEntityType = eventEntity.get("type")
self.setFeature("rootType_" + eventEntity.get("type"), 1)
self.setFeature("predStrength" + eventEntityType, self.getPredictionStrength(eventEntity))
self.triggerFeatureBuilder.setFeatureVector(self.features)
self.triggerFeatureBuilder.tag = "trg" + eventEntityType + "_"
self.triggerFeatureBuilder.buildFeatures(token)
self.triggerFeatureBuilder.tag = None
argThemeCount = 0
argCauseCount = 0
# Current example's edge combination
for i in range(len(interactions)):
arg = interactions[i]
if arg.get("type") == "Theme":
argThemeCount += 1
self.buildArgumentFeatures(sentenceGraph, paths, self.features, token, arg, "argTheme")
self.buildArgumentFeatures(sentenceGraph, paths, self.features, token, arg, "argTheme" + str(i))
else: # Cause
argCauseCount += 1
self.buildArgumentFeatures(sentenceGraph, paths, self.features, token, arg, "argCause")
self.buildArgumentFeatures(sentenceGraph, paths, self.features, token, arg, "argCause" + str(i))
self.setFeature("argCount", len(interactions))
self.setFeature("argCount_" + str(len(interactions)), 1)
self.setFeature("argThemeCount", argThemeCount)
self.setFeature("argThemeCount_" + str(argThemeCount), 1)
self.setFeature("argCauseCount", argCauseCount)
self.setFeature("argCauseCount_" + str(argCauseCount), 1)
self.triggerFeatureBuilder.tag = ""
self.triggerFeatureBuilder.setFeatureVector(None)
def buildArgumentFeatures(self, sentenceGraph, paths, features, eventToken, arg, tag):
argEntity = sentenceGraph.entitiesById[arg.get("e2")]
argToken = sentenceGraph.entityHeadTokenByEntity[argEntity]
self.buildEdgeFeatures(sentenceGraph, paths, features, eventToken, argToken, tag)
self.triggerFeatureBuilder.tag = tag + "trg_"
self.triggerFeatureBuilder.buildFeatures(argToken)
if argEntity.get("isName") == "True":
self.setFeature(tag + "Protein", 1)
else:
self.setFeature(tag + "Event", 1)
self.setFeature("nestingEvent", 1)
self.setFeature(tag + "_" + argEntity.get("type"), 1)
def buildEdgeFeatures(self, sentenceGraph, paths, features, eventToken, argToken, tag):
# eventToken = sentenceGraph.entityHeadTokenByEntity[eventNode]
# argToken = sentenceGraph.entityHeadTokenByEntity[argNode]
self.multiEdgeFeatureBuilder.tag = tag + "_"
self.multiEdgeFeatureBuilder.setFeatureVector(features, None, None, False)
self.setFeature(tag + "_present", 1)
if eventToken != argToken and paths.has_key(eventToken) and paths[eventToken].has_key(argToken):
path = paths[eventToken][argToken]
edges = self.multiEdgeFeatureBuilder.getEdges(sentenceGraph.dependencyGraph, path)
else:
path = [eventToken, argToken]
edges = None
if not "disable_entity_features" in self.styles:
self.multiEdgeFeatureBuilder.buildEntityFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.buildPathLengthFeatures(path)
# if not "disable_terminus_features" in self.styles:
# self.multiEdgeFeatureBuilder.buildTerminusTokenFeatures(path, sentenceGraph) # remove for fast
if not "disable_single_element_features" in self.styles:
self.multiEdgeFeatureBuilder.buildSingleElementFeatures(path, edges, sentenceGraph)
if not "disable_ngram_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathGrams(2, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(3, path, edges, sentenceGraph) # remove for fast
self.multiEdgeFeatureBuilder.buildPathGrams(4, path, edges, sentenceGraph) # remove for fast
if not "disable_path_edge_features" in self.styles:
self.multiEdgeFeatureBuilder.buildPathEdgeFeatures(path, edges, sentenceGraph)
# self.multiEdgeFeatureBuilder.buildSentenceFeatures(sentenceGraph)
self.multiEdgeFeatureBuilder.setFeatureVector(None, None, None, False)
self.multiEdgeFeatureBuilder.tag = ""
def buildInterArgumentBagOfWords(self, arguments, sentenceGraph):
if len(arguments) < 2:
return
indexByToken = {}
for i in range(len(sentenceGraph.tokens)):
indexByToken[sentenceGraph.tokens[i]] = i
argTokenIndices = set()
for arg in arguments:
argEntity = sentenceGraph.entitiesById[arg.get("e2")]
argToken = sentenceGraph.entityHeadTokenByEntity[argEntity]
argTokenIndices.add(indexByToken[argToken])
minIndex = min(argTokenIndices)
maxIndex = max(argTokenIndices)
self.setFeature("argBoWRange", (maxIndex - minIndex))
self.setFeature("argBoWRange_" + str(maxIndex - minIndex), 1)
bow = set()
for i in range(minIndex + 1, maxIndex):
token = sentenceGraph.tokens[i]
if len(sentenceGraph.tokenIsEntityHead[token]) == 0 and not sentenceGraph.tokenIsName[token]:
bow.add(token.get("text"))
bow = sorted(list(bow))
for word in bow:
self.setFeature("argBoW_" + word, 1)
if word in ["/", "-"]:
self.setFeature("argBoW_slashOrHyphen", 1)
if len(bow) == 1:
self.setFeature("argBoWonly_" + bow[0], 1)
if bow[0] in ["/", "-"]:
self.setFeature("argBoWonly_slashOrHyphen", 1)