def buildExamplesFromGraph(self, sentenceGraph, outfile, goldGraph = None):
"""
Build examples for a single sentence. Returns a list of examples.
See Core/ExampleUtils for example format.
"""
#examples = []
exampleIndex = 0
if self.styles["trigger_features"]:
self.triggerFeatureBuilder.initSentence(sentenceGraph)
if self.styles["evex"]:
self.evexFeatureBuilder.initSentence(sentenceGraph)
# Filter entities, if needed
#mergedIds = None
#duplicateEntities = None
#entities = sentenceGraph.entities
#entities, mergedIds, duplicateEntities = self.mergeEntities(sentenceGraph, False) # "no_duplicates" in self.styles)
sentenceGraph.mergeInteractionGraph(True)
entities = sentenceGraph.mergedEntities
entityToDuplicates = sentenceGraph.mergedEntityToDuplicates
self.exampleStats.addValue("Duplicate entities skipped", len(sentenceGraph.entities) - len(entities))
# Connect to optional gold graph
if goldGraph != None:
entityToGold = EvaluateInteractionXML.mapEntities(entities, goldGraph.entities)
paths = None
if not self.styles["no_path"]:
##undirected = sentenceGraph.getUndirectedDependencyGraph()
#undirected = self.nxMultiDiGraphToUndirected(sentenceGraph.dependencyGraph)
###undirected = sentenceGraph.dependencyGraph.to_undirected()
####undirected = NX10.MultiGraph(sentenceGraph.dependencyGraph) This didn't work
undirected = sentenceGraph.dependencyGraph.toUndirected()
#paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
paths = undirected
#for edge in sentenceGraph.dependencyGraph.edges:
# assert edge[2] != None
#for edge in undirected.edges:
# assert edge[2] != None
#if sentenceGraph.sentenceElement.get("id") == "GENIA.d70.s5":
# print [(x[0].get("id"), x[1].get("id"), x[2].get("id")) for x in sentenceGraph.dependencyGraph.edges]
# Generate examples based on interactions between entities or interactions between tokens
if self.styles["entities"]:
loopRange = len(entities)
else:
loopRange = len(sentenceGraph.tokens)
for i in range(loopRange-1):
for j in range(i+1,loopRange):
eI = None
eJ = None
if self.styles["entities"]:
eI = entities[i]
eJ = 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
if self.styles["skip_extra_triggers"]:
if eI.get("source") != None or eJ.get("source") != None:
continue
else:
tI = sentenceGraph.tokens[i]
tJ = sentenceGraph.tokens[j]
# only consider paths between entities (NOTE! entities, not only named entities)
if self.styles["headsOnly"]:
if (len(sentenceGraph.tokenIsEntityHead[tI]) == 0) or (len(sentenceGraph.tokenIsEntityHead[tJ]) == 0):
continue
if self.styles["directed"]:
# define forward
if self.styles["entities"]:
categoryName = self.getCategoryName(sentenceGraph, eI, eJ, True)
if goldGraph != None:
categoryName = self.getGoldCategoryName(goldGraph, entityToGold, eI, eJ, True)
else:
categoryName = self.getCategoryNameFromTokens(sentenceGraph, tI, tJ, True)
# make forward
self.exampleStats.beginExample(categoryName)
makeExample = True
if self.styles["genia_limits"] and not self.isPotentialGeniaInteraction(eI, eJ):
makeExample = False
self.exampleStats.filter("genia_limits")
if self.styles["genia_task1"] and (eI.get("type") == "Entity" or eJ.get("type") == "Entity"):
makeExample = False
self.exampleStats.filter("genia_task1")
if self.styles["rel_limits"] and not self.isPotentialRELInteraction(eI, eJ):
makeExample = False
self.exampleStats.filter("rel_limits")
if self.styles["co_limits"] and not self.isPotentialCOInteraction(eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("co_limits")
if self.styles["bb_limits"] and not self.isPotentialBBInteraction(eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("bb_limits")
if categoryName != "neg":
self.exampleStats.filter("bb_limits(" + categoryName + ":" + eI.get("type") + "/" + eJ.get("type") + ")")
if self.styles["bi_limits"] and not self.isPotentialBIInteraction(eI, eJ, sentenceGraph, self.exampleStats):
makeExample = False
#self.exampleStats.filter("bi_limits")
if self.styles["epi_limits"] and not self.isPotentialEPIInteraction(eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("epi_limits")
if self.styles["id_limits"] and not self.isPotentialIDInteraction(eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("id_limits")
# if self.styles["selftrain_limits"] and (eI.get("selftrain") == "False" or eJ.get("selftrain") == "False"):
# makeExample = False
# self.exampleStats.filter("selftrain_limits")
# if self.styles["selftrain_group"] and (eI.get("selftraingroup") not in self.selfTrainGroups or eJ.get("selftraingroup") not in self.selfTrainGroups):
# makeExample = False
# self.exampleStats.filter("selftrain_group")
if self.styles["pos_only"] and categoryName == "neg":
makeExample = False
self.exampleStats.filter("pos_only")
if makeExample:
#examples.append( self.buildExample(tI, tJ, paths, sentenceGraph, categoryName, exampleIndex, eI, eJ) )
ExampleUtils.appendExamples([self.buildExample(tI, tJ, paths, sentenceGraph, categoryName, exampleIndex, eI, eJ)], outfile)
exampleIndex += 1
self.exampleStats.endExample()
# define reverse
if self.styles["entities"]:
categoryName = self.getCategoryName(sentenceGraph, eJ, eI, True)
if goldGraph != None:
categoryName = self.getGoldCategoryName(goldGraph, entityToGold, eJ, eI, True)
else:
categoryName = self.getCategoryNameFromTokens(sentenceGraph, tJ, tI, True)
# make reverse
self.exampleStats.beginExample(categoryName)
makeExample = True
if self.styles["genia_limits"] and not self.isPotentialGeniaInteraction(eJ, eI):
makeExample = False
self.exampleStats.filter("genia_limits")
if self.styles["genia_task1"] and (eI.get("type") == "Entity" or eJ.get("type") == "Entity"):
makeExample = False
self.exampleStats.filter("genia_task1")
if self.styles["rel_limits"] and not self.isPotentialRELInteraction(eJ, eI):
makeExample = False
self.exampleStats.filter("rel_limits")
if self.styles["co_limits"] and not self.isPotentialCOInteraction(eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("co_limits")
if self.styles["bb_limits"] and not self.isPotentialBBInteraction(eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("bb_limits")
if categoryName != "neg":
self.exampleStats.filter("bb_limits(" + categoryName + ":" + eJ.get("type") + "/" + eI.get("type") + ")")
if self.styles["bi_limits"] and not self.isPotentialBIInteraction(eJ, eI, sentenceGraph, self.exampleStats):
makeExample = False
#self.exampleStats.filter("bi_limits")
if self.styles["epi_limits"] and not self.isPotentialEPIInteraction(eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("epi_limits")
if self.styles["id_limits"] and not self.isPotentialIDInteraction(eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("id_limits")
# if self.styles["selftrain_limits"] and (eI.get("selftrain") == "False" or eJ.get("selftrain") == "False"):
# makeExample = False
# self.exampleStats.filter("selftrain_limits")
# if self.styles["selftrain_group"] and (eI.get("selftraingroup") not in self.selfTrainGroups or eJ.get("selftraingroup") not in self.selfTrainGroups):
# makeExample = False
# self.exampleStats.filter("selftrain_group")
if self.styles["pos_only"] and categoryName == "neg":
makeExample = False
self.exampleStats.filter("pos_only")
if makeExample:
#examples.append( self.buildExample(tJ, tI, paths, sentenceGraph, categoryName, exampleIndex, eJ, eI) )
ExampleUtils.appendExamples([self.buildExample(tJ, tI, paths, sentenceGraph, categoryName, exampleIndex, eJ, eI)], outfile)
exampleIndex += 1
self.exampleStats.endExample()
else:
if self.styles["entities"]:
categoryName = self.getCategoryName(sentenceGraph, eI, eJ, False)
else:
categoryName = self.getCategoryNameFromTokens(sentenceGraph, tI, tJ, False)
self.exampleStats.beginExample(categoryName)
forwardExample = self.buildExample(tI, tJ, paths, sentenceGraph, categoryName, exampleIndex, eI, eJ)
if not self.styles["graph_kernel"]:
reverseExample = self.buildExample(tJ, tI, paths, sentenceGraph, categoryName, exampleIndex, eJ, eI)
forwardExample[2].update(reverseExample[2])
#examples.append(forwardExample)
ExampleUtils.appendExamples([forwardExample], outfile)
exampleIndex += 1
self.exampleStats.endExample()
#return examples
return exampleIndex
def buildExamplesFromGraph(self, sentenceGraph, examples, goldGraph=None):
# example directionality
if self.styles.get("directed") == None and self.styles.get("undirected") == None: # determine directedness from corpus
examplesAreDirected = self.structureAnalyzer.hasDirectedTargets() if self.structureAnalyzer != None else True
elif self.styles.get("directed"):
assert self.styles.get("undirected") in [None, False]
examplesAreDirected = True
elif self.styles.get("undirected"):
assert self.styles.get("directed") in [None, False]
examplesAreDirected = False
# Filter entities, if needed
sentenceGraph.mergeInteractionGraph(True)
entities = sentenceGraph.mergedEntities
#entityToDuplicates = sentenceGraph.mergedEntityToDuplicates
self.exampleStats.addValue("Duplicate entities skipped", len(sentenceGraph.entities) - len(entities))
# Connect to optional gold graph
entityToGold = None
if goldGraph != None:
entityToGold = EvaluateInteractionXML.mapEntities(entities, goldGraph.entities)
# paths = None
# if not self.styles.get("no_path"):
# undirected = sentenceGraph.dependencyGraph.toUndirected()
# paths = undirected
# if self.styles.get("filter_shortest_path") != None: # For DDI use filter_shortest_path=conj_and
# paths.resetAnalyses() # just in case
# paths.FloydWarshall(self.filterEdge, {"edgeTypes":self.styles["filter_shortest_path"]})
dg = sentenceGraph.dependencyGraph
undirected = dg.toUndirected()
edgeCounts = {x:len(dg.getInEdges(x) + dg.getOutEdges(x)) for x in sentenceGraph.tokens}
tokens, tokenMap = self.getTokenFeatures(sentenceGraph)
# Generate examples based on interactions between entities or interactions between tokens
if self.styles.get("token_nodes"):
loopRange = len(tokens)
else:
loopRange = len(entities)
for i in range(loopRange-1):
for j in range(i+1,loopRange):
eI = None
eJ = None
if self.styles.get("token_nodes"):
tI = tokens[i]["element"]
tJ = tokens[j]["element"]
else:
eI = entities[i]
eJ = entities[j]
tI = sentenceGraph.entityHeadTokenByEntity[eI]
tJ = sentenceGraph.entityHeadTokenByEntity[eJ]
if eI.get("type") == "neg" or eJ.get("type") == "neg":
continue
if self.styles.get("skip_extra_triggers"):
if eI.get("source") != None or eJ.get("source") != None:
continue
# only consider paths between entities (NOTE! entities, not only named entities)
if self.styles.get("headsOnly"):
if (len(sentenceGraph.tokenIsEntityHead[tI]) == 0) or (len(sentenceGraph.tokenIsEntityHead[tJ]) == 0):
continue
if examplesAreDirected:
self.buildExample(examples, tI, tJ, eI, eJ, tokens, tokenMap, sentenceGraph, goldGraph, entityToGold, undirected, edgeCounts)
self.buildExample(examples, tJ, tI, eJ, eI, tokens, tokenMap, sentenceGraph, goldGraph, entityToGold, undirected, edgeCounts)
else:
if tokenMap[tJ]["index"] < tokenMap[tI]["index"]:
tI, tJ = tJ, tI
eI, eJ = eJ, eI
self.buildExample(examples, tI, tJ, eI, eJ, tokens, tokenMap, sentenceGraph, goldGraph, entityToGold, undirected, edgeCounts, False)
def buildExamplesFromGraph(self, sentenceGraph, outfile, goldGraph = None, structureAnalyzer=None):
"""
Build examples for a single sentence. Returns a list of examples.
See Core/ExampleUtils for example format.
"""
#examples = []
exampleIndex = 0
# example directionality
if self.styles["directed"] == None and self.styles["undirected"] == None: # determine directedness from corpus
examplesAreDirected = structureAnalyzer.hasDirectedTargets() if structureAnalyzer != None else True
elif self.styles["directed"]:
assert self.styles["undirected"] in [None, False]
examplesAreDirected = True
elif self.styles["undirected"]:
assert self.styles["directed"] in [None, False]
examplesAreDirected = False
if not self.styles["no_trigger_features"]:
self.triggerFeatureBuilder.initSentence(sentenceGraph)
if self.styles["evex"]:
self.evexFeatureBuilder.initSentence(sentenceGraph)
# if self.styles["sdb_merge"]:
# self.determineNonOverlappingTypes(structureAnalyzer)
# Filter entities, if needed
sentenceGraph.mergeInteractionGraph(True)
entities = sentenceGraph.mergedEntities
entityToDuplicates = sentenceGraph.mergedEntityToDuplicates
self.exampleStats.addValue("Duplicate entities skipped", len(sentenceGraph.entities) - len(entities))
# Connect to optional gold graph
entityToGold = None
if goldGraph != None:
entityToGold = EvaluateInteractionXML.mapEntities(entities, goldGraph.entities)
paths = None
if not self.styles["no_path"]:
undirected = sentenceGraph.dependencyGraph.toUndirected()
paths = undirected
if self.styles["filter_shortest_path"] != None: # For DDI use filter_shortest_path=conj_and
paths.resetAnalyses() # just in case
paths.FloydWarshall(self.filterEdge, {"edgeTypes":self.styles["filter_shortest_path"]})
# Generate examples based on interactions between entities or interactions between tokens
if self.styles["token_nodes"]:
loopRange = len(sentenceGraph.tokens)
else:
loopRange = len(entities)
for i in range(loopRange-1):
for j in range(i+1,loopRange):
eI = None
eJ = None
if self.styles["token_nodes"]:
tI = sentenceGraph.tokens[i]
tJ = sentenceGraph.tokens[j]
else:
eI = entities[i]
eJ = entities[j]
tI = sentenceGraph.entityHeadTokenByEntity[eI]
tJ = sentenceGraph.entityHeadTokenByEntity[eJ]
if eI.get("type") == "neg" or eJ.get("type") == "neg":
continue
if self.styles["skip_extra_triggers"]:
if eI.get("source") != None or eJ.get("source") != None:
continue
# only consider paths between entities (NOTE! entities, not only named entities)
if self.styles["headsOnly"]:
if (len(sentenceGraph.tokenIsEntityHead[tI]) == 0) or (len(sentenceGraph.tokenIsEntityHead[tJ]) == 0):
continue
examples = self.buildExamplesForPair(tI, tJ, paths, sentenceGraph, goldGraph, entityToGold, eI, eJ, structureAnalyzer, examplesAreDirected)
for categoryName, features, extra in examples:
# make example
if self.styles["binary"]:
if categoryName != "neg":
category = 1
else:
category = -1
extra["categoryName"] = "i"
else:
category = self.classSet.getId(categoryName)
example = [sentenceGraph.getSentenceId()+".x"+str(exampleIndex), category, features, extra]
ExampleUtils.appendExamples([example], outfile)
exampleIndex += 1
return exampleIndex
def buildExamplesFromGraph(self, sentenceGraph, outfile, goldGraph=None, structureAnalyzer=None):
"""
Build one example for each token of the sentence
"""
examples = []
exampleIndex = 0
self.tokenFeatures = {}
if goldGraph != None:
entityToGold = EvaluateInteractionXML.mapEntities(sentenceGraph.entities, goldGraph.entities)
namedEntityCount = 0
entityCount = 0
for entity in sentenceGraph.entities:
if entity.get("given") == "True": # known data which can be used for features
namedEntityCount += 1
else: # known data which can be used for features
entityCount += 1
namedEntityCountFeature = "nameCount_" + str(namedEntityCount)
entityCountFeature = "entityCount_" + str(entityCount)
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
if len(sentenceGraph.tokenIsEntityHead) > 0:
text = "ge_" + text
if not bagOfWords.has_key(text):
bagOfWords[text] = 0
bagOfWords[text] += 1
text = token.get("text")
if self.styles["speculation_words"] and text in self.specWords:
if not bagOfWords.has_key("spec_bow_"+text):
bagOfWords["spec_bow_"+text] = 0
bagOfWords["spec_bow_"+text] += 1
bagOfWords["spec_sentence"] = 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.getInEdges(token)
self.inEdgesByToken[token] = inEdges
outEdges = sentenceGraph.dependencyGraph.getOutEdges(token)
self.outEdgesByToken[token] = outEdges
self.edgeSetByToken[token] = set(inEdges + outEdges)
for entity in sentenceGraph.entities:
#token = sentenceGraph.tokens[i]
token = sentenceGraph.entityHeadTokenByEntity[entity]
# Recognize only non-named entities (i.e. interaction words)
if entity.get("given") == "True":
continue
# CLASS
if self.styles["classification"] == "multiclass":
task3Type = "multiclass"
categoryName = ""
if entity.get("negation") == "True":
categoryName += "negation"
if entity.get("speculation") == "True":
if categoryName != "":
categoryName += "---"
categoryName += "speculation"
if categoryName == "":
categoryName = "neg"
category = self.classSet.getId(categoryName)
elif self.styles["classification"] == "speculation":
task3Type = "speculation"
if entity.get("speculation") == "True":
category = self.classSet.getId("speculation")
else:
category = 1
if goldGraph != None:
if len(entityToGold[entity]) > 0 and entityToGold[entity][0].get("speculation") == "True":
category = self.classSet.getId("speculation")
else:
category = 1
categoryName = self.classSet.getName(category)
elif self.styles["classification"] == "negation":
task3Type = "negation"
if entity.get("negation") == "True":
category = self.classSet.getId("negation")
else:
category = 1
if goldGraph != None:
if len(entityToGold[entity]) > 0 and entityToGold[entity][0].get("negation") == "True":
category = self.classSet.getId("negation")
else:
category = 1
categoryName = self.classSet.getName(category)
self.exampleStats.beginExample(categoryName)
# FEATURES
features = {}
# ENTITY TYPE
#entityType = self.classSet.getId(self.getMergedEntityType(entity))
#del self.classSet.Ids[self.getMergedEntityType(entity)]
#IF LOCAL
# There's a mistake here. The entityType should be the string, not
# the id of the type. But there's also another issue. getMergedEntityType
# expects a list, not an item. Therefore the type is always empty ->
# types don't get used in classification. But this is the code used in
# the publication, so it will now be published as is, and fixed in a later
# release.
#
# Besides, using the classSet here generates an unneeded
# additional class, that shows up in evaluations etc. However, to be
# able to publish the exact models used for the publication experiments,
# this can't be fixed so it breaks feature id consistency. Therefore I'll
# now just remove the redundant class id from the classSet.
#ENDIF
#features[self.featureSet.getId(entityType)] = 1
features[self.featureSet.getId(namedEntityCountFeature)] = 1
features[self.featureSet.getId(entityCountFeature)] = 1
#for k,v in bagOfWords.iteritems():
# features[self.featureSet.getId(k)] = v
# pre-calculate bow _features_
features.update(bowFeatures)
# 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
if self.styles["speculation_words"]:
if text in self.specWords:
features[self.featureSet.getId("ent_spec")] = 1
if stem in self.specWordStems:
features[self.featureSet.getId("ent_spec_stem")] = 1
# Linear order features
for i in range(len(sentenceGraph.tokens)):
if token == sentenceGraph.tokens[i]:
break
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
self.buildChains(token, sentenceGraph, features)
extra = {"xtype":"task3","t3type":task3Type,"t":token.get("id"),"entity":entity.get("id")}
#examples.append( (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) )
example = (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra)
ExampleUtils.appendExamples([example], outfile)
exampleIndex += 1
self.exampleStats.endExample()
#return examples
return exampleIndex
def buildExamplesFromGraph(self, sentenceGraph, outfile, goldGraph=None):
"""
Build examples for a single sentence. Returns a list of examples.
See Core/ExampleUtils for example format.
"""
#examples = []
exampleIndex = 0
if self.styles["trigger_features"]:
self.triggerFeatureBuilder.initSentence(sentenceGraph)
if self.styles["evex"]:
self.evexFeatureBuilder.initSentence(sentenceGraph)
# Filter entities, if needed
#mergedIds = None
#duplicateEntities = None
#entities = sentenceGraph.entities
#entities, mergedIds, duplicateEntities = self.mergeEntities(sentenceGraph, False) # "no_duplicates" in self.styles)
sentenceGraph.mergeInteractionGraph(True)
entities = sentenceGraph.mergedEntities
entityToDuplicates = sentenceGraph.mergedEntityToDuplicates
self.exampleStats.addValue("Duplicate entities skipped",
len(sentenceGraph.entities) - len(entities))
# Connect to optional gold graph
if goldGraph != None:
entityToGold = EvaluateInteractionXML.mapEntities(
entities, goldGraph.entities)
paths = None
if not self.styles["no_path"]:
##undirected = sentenceGraph.getUndirectedDependencyGraph()
#undirected = self.nxMultiDiGraphToUndirected(sentenceGraph.dependencyGraph)
###undirected = sentenceGraph.dependencyGraph.to_undirected()
####undirected = NX10.MultiGraph(sentenceGraph.dependencyGraph) This didn't work
undirected = sentenceGraph.dependencyGraph.toUndirected()
#paths = NX10.all_pairs_shortest_path(undirected, cutoff=999)
paths = undirected
#for edge in sentenceGraph.dependencyGraph.edges:
# assert edge[2] != None
#for edge in undirected.edges:
# assert edge[2] != None
#if sentenceGraph.sentenceElement.get("id") == "GENIA.d70.s5":
# print [(x[0].get("id"), x[1].get("id"), x[2].get("id")) for x in sentenceGraph.dependencyGraph.edges]
# Generate examples based on interactions between entities or interactions between tokens
if self.styles["entities"]:
loopRange = len(entities)
else:
loopRange = len(sentenceGraph.tokens)
for i in range(loopRange - 1):
for j in range(i + 1, loopRange):
eI = None
eJ = None
if self.styles["entities"]:
eI = entities[i]
eJ = 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
if self.styles["skip_extra_triggers"]:
if eI.get("source") != None or eJ.get(
"source") != None:
continue
else:
tI = sentenceGraph.tokens[i]
tJ = sentenceGraph.tokens[j]
# only consider paths between entities (NOTE! entities, not only named entities)
if self.styles["headsOnly"]:
if (len(sentenceGraph.tokenIsEntityHead[tI]) == 0) or (len(
sentenceGraph.tokenIsEntityHead[tJ]) == 0):
continue
if self.styles["directed"]:
# define forward
if self.styles["entities"]:
categoryName = self.getCategoryName(
sentenceGraph, eI, eJ, True)
if goldGraph != None:
categoryName = self.getGoldCategoryName(
goldGraph, entityToGold, eI, eJ, True)
else:
categoryName = self.getCategoryNameFromTokens(
sentenceGraph, tI, tJ, True)
# make forward
self.exampleStats.beginExample(categoryName)
makeExample = True
if self.styles[
"genia_limits"] and not self.isPotentialGeniaInteraction(
eI, eJ):
makeExample = False
self.exampleStats.filter("genia_limits")
if self.styles["genia_task1"] and (
eI.get("type") == "Entity"
or eJ.get("type") == "Entity"):
makeExample = False
self.exampleStats.filter("genia_task1")
if self.styles[
"rel_limits"] and not self.isPotentialRELInteraction(
eI, eJ):
makeExample = False
self.exampleStats.filter("rel_limits")
if self.styles[
"co_limits"] and not self.isPotentialCOInteraction(
eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("co_limits")
if self.styles[
"bb_limits"] and not self.isPotentialBBInteraction(
eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("bb_limits")
if categoryName != "neg":
self.exampleStats.filter("bb_limits(" +
categoryName + ":" +
eI.get("type") + "/" +
eJ.get("type") + ")")
if self.styles[
"bi_limits"] and not self.isPotentialBIInteraction(
eI, eJ, sentenceGraph, self.exampleStats):
makeExample = False
#self.exampleStats.filter("bi_limits")
if self.styles[
"epi_limits"] and not self.isPotentialEPIInteraction(
eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("epi_limits")
if self.styles[
"id_limits"] and not self.isPotentialIDInteraction(
eI, eJ, sentenceGraph):
makeExample = False
self.exampleStats.filter("id_limits")
# if self.styles["selftrain_limits"] and (eI.get("selftrain") == "False" or eJ.get("selftrain") == "False"):
# makeExample = False
# self.exampleStats.filter("selftrain_limits")
# if self.styles["selftrain_group"] and (eI.get("selftraingroup") not in self.selfTrainGroups or eJ.get("selftraingroup") not in self.selfTrainGroups):
# makeExample = False
# self.exampleStats.filter("selftrain_group")
if self.styles["pos_only"] and categoryName == "neg":
makeExample = False
self.exampleStats.filter("pos_only")
if makeExample:
#examples.append( self.buildExample(tI, tJ, paths, sentenceGraph, categoryName, exampleIndex, eI, eJ) )
ExampleUtils.appendExamples([
self.buildExample(tI, tJ, paths, sentenceGraph,
categoryName, exampleIndex, eI,
eJ)
], outfile)
exampleIndex += 1
self.exampleStats.endExample()
# define reverse
if self.styles["entities"]:
categoryName = self.getCategoryName(
sentenceGraph, eJ, eI, True)
if goldGraph != None:
categoryName = self.getGoldCategoryName(
goldGraph, entityToGold, eJ, eI, True)
else:
categoryName = self.getCategoryNameFromTokens(
sentenceGraph, tJ, tI, True)
# make reverse
self.exampleStats.beginExample(categoryName)
makeExample = True
if self.styles[
"genia_limits"] and not self.isPotentialGeniaInteraction(
eJ, eI):
makeExample = False
self.exampleStats.filter("genia_limits")
if self.styles["genia_task1"] and (
eI.get("type") == "Entity"
or eJ.get("type") == "Entity"):
makeExample = False
self.exampleStats.filter("genia_task1")
if self.styles[
"rel_limits"] and not self.isPotentialRELInteraction(
eJ, eI):
makeExample = False
self.exampleStats.filter("rel_limits")
if self.styles[
"co_limits"] and not self.isPotentialCOInteraction(
eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("co_limits")
if self.styles[
"bb_limits"] and not self.isPotentialBBInteraction(
eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("bb_limits")
if categoryName != "neg":
self.exampleStats.filter("bb_limits(" +
categoryName + ":" +
eJ.get("type") + "/" +
eI.get("type") + ")")
if self.styles[
"bi_limits"] and not self.isPotentialBIInteraction(
eJ, eI, sentenceGraph, self.exampleStats):
makeExample = False
#self.exampleStats.filter("bi_limits")
if self.styles[
"epi_limits"] and not self.isPotentialEPIInteraction(
eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("epi_limits")
if self.styles[
"id_limits"] and not self.isPotentialIDInteraction(
eJ, eI, sentenceGraph):
makeExample = False
self.exampleStats.filter("id_limits")
# if self.styles["selftrain_limits"] and (eI.get("selftrain") == "False" or eJ.get("selftrain") == "False"):
# makeExample = False
# self.exampleStats.filter("selftrain_limits")
# if self.styles["selftrain_group"] and (eI.get("selftraingroup") not in self.selfTrainGroups or eJ.get("selftraingroup") not in self.selfTrainGroups):
# makeExample = False
# self.exampleStats.filter("selftrain_group")
if self.styles["pos_only"] and categoryName == "neg":
makeExample = False
self.exampleStats.filter("pos_only")
if makeExample:
#examples.append( self.buildExample(tJ, tI, paths, sentenceGraph, categoryName, exampleIndex, eJ, eI) )
ExampleUtils.appendExamples([
self.buildExample(tJ, tI, paths, sentenceGraph,
categoryName, exampleIndex, eJ,
eI)
], outfile)
exampleIndex += 1
self.exampleStats.endExample()
else:
if self.styles["entities"]:
categoryName = self.getCategoryName(
sentenceGraph, eI, eJ, False)
else:
categoryName = self.getCategoryNameFromTokens(
sentenceGraph, tI, tJ, False)
self.exampleStats.beginExample(categoryName)
forwardExample = self.buildExample(tI, tJ, paths,
sentenceGraph,
categoryName,
exampleIndex, eI, eJ)
if not self.styles["graph_kernel"]:
reverseExample = self.buildExample(
tJ, tI, paths, sentenceGraph, categoryName,
exampleIndex, eJ, eI)
forwardExample[2].update(reverseExample[2])
#examples.append(forwardExample)
ExampleUtils.appendExamples([forwardExample], outfile)
exampleIndex += 1
self.exampleStats.endExample()
#return examples
return exampleIndex
def buildExamplesFromGraph(self, sentenceGraph, outfile, goldGraph=None):
"""
Build one example for each token of the sentence
"""
examples = []
exampleIndex = 0
self.tokenFeatures = {}
if goldGraph != None:
entityToGold = EvaluateInteractionXML.mapEntities(
sentenceGraph.entities, goldGraph.entities)
namedEntityCount = 0
entityCount = 0
for entity in sentenceGraph.entities:
if entity.get(
"isName"
) == "True": # known data which can be used for features
namedEntityCount += 1
else: # known data which can be used for features
entityCount += 1
namedEntityCountFeature = "nameCount_" + str(namedEntityCount)
entityCountFeature = "entityCount_" + str(entityCount)
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
if len(sentenceGraph.tokenIsEntityHead) > 0:
text = "ge_" + text
if not bagOfWords.has_key(text):
bagOfWords[text] = 0
bagOfWords[text] += 1
text = token.get("text")
if self.styles["speculation_words"] and text in self.specWords:
if not bagOfWords.has_key("spec_bow_" + text):
bagOfWords["spec_bow_" + text] = 0
bagOfWords["spec_bow_" + text] += 1
bagOfWords["spec_sentence"] = 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.getInEdges(token)
self.inEdgesByToken[token] = inEdges
outEdges = sentenceGraph.dependencyGraph.getOutEdges(token)
self.outEdgesByToken[token] = outEdges
self.edgeSetByToken[token] = set(inEdges + outEdges)
for entity in sentenceGraph.entities:
#token = sentenceGraph.tokens[i]
token = sentenceGraph.entityHeadTokenByEntity[entity]
# Recognize only non-named entities (i.e. interaction words)
if entity.get("isName") == "True":
continue
# CLASS
if self.styles["classification"] == "multiclass":
task3Type = "multiclass"
categoryName = ""
if entity.get("negation") == "True":
categoryName += "negation"
if entity.get("speculation") == "True":
if categoryName != "":
categoryName += "---"
categoryName += "speculation"
if categoryName == "":
categoryName = "neg"
category = self.classSet.getId(categoryName)
elif self.styles["classification"] == "speculation":
task3Type = "speculation"
if entity.get("speculation") == "True":
category = self.classSet.getId("speculation")
else:
category = 1
if goldGraph != None:
if len(entityToGold[entity]) > 0 and entityToGold[entity][
0].get("speculation") == "True":
category = self.classSet.getId("speculation")
else:
category = 1
categoryName = self.classSet.getName(category)
elif self.styles["classification"] == "negation":
task3Type = "negation"
if entity.get("negation") == "True":
category = self.classSet.getId("negation")
else:
category = 1
if goldGraph != None:
if len(entityToGold[entity]) > 0 and entityToGold[entity][
0].get("negation") == "True":
category = self.classSet.getId("negation")
else:
category = 1
categoryName = self.classSet.getName(category)
self.exampleStats.beginExample(categoryName)
# FEATURES
features = {}
# ENTITY TYPE
#entityType = self.classSet.getId(self.getMergedEntityType(entity))
#del self.classSet.Ids[self.getMergedEntityType(entity)]
#IF LOCAL
# There's a mistake here. The entityType should be the string, not
# the id of the type. But there's also another issue. getMergedEntityType
# expects a list, not an item. Therefore the type is always empty ->
# types don't get used in classification. But this is the code used in
# the publication, so it will now be published as is, and fixed in a later
# release.
#
# Besides, using the classSet here generates an unneeded
# additional class, that shows up in evaluations etc. However, to be
# able to publish the exact models used for the publication experiments,
# this can't be fixed so it breaks feature id consistency. Therefore I'll
# now just remove the redundant class id from the classSet.
#ENDIF
#features[self.featureSet.getId(entityType)] = 1
features[self.featureSet.getId(namedEntityCountFeature)] = 1
features[self.featureSet.getId(entityCountFeature)] = 1
#for k,v in bagOfWords.iteritems():
# features[self.featureSet.getId(k)] = v
# pre-calculate bow _features_
features.update(bowFeatures)
# 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
if self.styles["speculation_words"]:
if text in self.specWords:
features[self.featureSet.getId("ent_spec")] = 1
if stem in self.specWordStems:
features[self.featureSet.getId("ent_spec_stem")] = 1
# Linear order features
for i in range(len(sentenceGraph.tokens)):
if token == sentenceGraph.tokens[i]:
break
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
self.buildChains(token, sentenceGraph, features)
extra = {
"xtype": "task3",
"t3type": task3Type,
"t": token.get("id"),
"entity": entity.get("id")
}
#examples.append( (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) )
example = (sentenceGraph.getSentenceId() + ".x" +
str(exampleIndex), category, features, extra)
ExampleUtils.appendExamples([example], outfile)
exampleIndex += 1
self.exampleStats.endExample()
#return examples
return exampleIndex
def makeExampleGraphWithGold(self, builder, sentenceGraph, goldGraph, sentenceIndex):
exampleGraph = NX10.MultiDiGraph()
for token in goldGraph.tokens:
exampleGraph.add_node(token)
arcStyles = {}
labelStyles = {}
extraByToken = {}
edgeTypes = {}
stats = {"entities":0,"edges":0,"tp":0,"fp":0,"tn":0,"fn":0}
entityMap = EvaluateInteractionXML.mapEntities(sentenceGraph.entities, goldGraph.entities, goldGraph.tokens)
tokenMap = self.getTokenMap(sentenceGraph, goldGraph)
toEntitiesWithPredictions = set()
for entityFrom, entitiesTo in entityMap.iteritems():
stats["entities"] += 1
entityFromHeadToken = sentenceGraph.entityHeadTokenByEntity[entityFrom]
for entityTo in entitiesTo:
toEntitiesWithPredictions.add(entityTo)
entityToHeadToken = goldGraph.entityHeadTokenByEntity[entityTo]
style = None
eFromType = entityFrom.get("type")
eToType = entityTo.get("type")
if extraByToken.has_key(entityToHeadToken):
style = extraByToken[entityToHeadToken]
if eFromType == eToType:
if eToType != "neg":
if style == None:
style = [entityTo.get("type"),{"fill":"green"}]
elif style[1]["fill"] == "#79BAEC":
style = [entityTo.get("type"),{"fill":"green"}]
if entityTo.get("isName") == "True":
style = [entityTo.get("type"),{"fill":"brown"}]
else:
stats["tp"] += 1
else:
if eToType == "neg":
pass
extraByToken[entityToHeadToken] = style
if len(entitiesTo) == 0:
stats["fp"] += 1
if extraByToken.has_key(tokenMap[entityFromHeadToken]):
style = extraByToken[tokenMap[entityFromHeadToken]]
if style[1]["fill"] != "green":
style = [entityFrom.get("type"),{"fill":"red"}]
extraByToken[tokenMap[entityFromHeadToken]] = style
else:
extraByToken[tokenMap[entityFromHeadToken]] = [entityFrom.get("type"),{"fill":"red"}]
for entity in goldGraph.entities:
if entity not in toEntitiesWithPredictions:
stats["fn"] += 1
extraByToken[goldGraph.entityHeadTokenByEntity[entity]] = [entity.get("type"),{"fill":"#79BAEC"}]
toInteractionsWithPredictions = set()
for interactionFrom in sentenceGraph.interactions:
if interactionFrom.get("type") == "neg":
continue
stats["edges"] += 1
e1s = entityMap[sentenceGraph.entitiesById[interactionFrom.get("e1")]]
e1Ids = []
for e1 in e1s:
e1Ids.append(e1.get("id"))
e2s = entityMap[sentenceGraph.entitiesById[interactionFrom.get("e2")]]
e2Ids = []
for e2 in e2s:
e2Ids.append(e2.get("id"))
t1 = tokenMap[sentenceGraph.entityHeadTokenByEntity[sentenceGraph.entitiesById[interactionFrom.get("e1")]]]
t2 = tokenMap[sentenceGraph.entityHeadTokenByEntity[sentenceGraph.entitiesById[interactionFrom.get("e2")]]]
iFromType = interactionFrom.get("type")
found = False
for interactionTo in goldGraph.interactions:
if interactionTo.get("e1") in e1Ids and interactionTo.get("e2") in e2Ids:
toInteractionsWithPredictions.add(interactionTo)
iToType = interactionTo.get("type")
exampleGraph.add_edge(t1, t2, element=interactionFrom)
#edge = exampleGraph.get_edge(t1, t2, data=True)
edge = self.getNXEdge(exampleGraph, t1, t2, interactionFrom)
if t1 != t2:
if iToType == iFromType:
edge[2]["arcStyles"] = {"stroke":"green"}
edge[2]["labelStyles"] = {"fill":"green"}
stats["tp"] += 1
else:
edge[2]["arcStyles"] = {"stroke":"red"}
edge[2]["labelStyles"] = {"fill":"red"}
stats["fp"] += 1
found = True
if not found: # false positive prediction
if t1 != t2:
exampleGraph.add_edge(t1, t2, element=interactionFrom)
edge = self.getNXEdge(exampleGraph, t1, t2, interactionFrom)
edge[2]["arcStyles"] = {"stroke":"red"}
edge[2]["labelStyles"] = {"fill":"red"}
stats["fp"] += 1
for interactionTo in goldGraph.interactions:
if interactionTo not in toInteractionsWithPredictions: # false negative gold
t1 = goldGraph.entityHeadTokenByEntity[goldGraph.entitiesById[interactionTo.get("e1")]]
t2 = goldGraph.entityHeadTokenByEntity[goldGraph.entitiesById[interactionTo.get("e2")]]
if t1 != t2:
exampleGraph.add_edge(t1, t2, element=interactionTo)
edge = self.getNXEdge(exampleGraph, t1, t2, interactionTo)
edge[2]["arcStyles"] = {"stroke":"#79BAEC"}
edge[2]["labelStyles"] = {"fill":"#79BAEC"}
stats["fn"] += 1
builder.header("Classification",4)
svgTokens = GraphToSVG.tokensToSVG(goldGraph.tokens,False,None,extraByToken)
#arcStyles, labelStyles = self.getMatchingEdgeStyles(exampleGraph, sentenceGraph.interactionGraph, "green", "red" )
svgEdges = GraphToSVG.edgesToSVG(svgTokens, exampleGraph, "type", None)
sentenceId = sentenceGraph.getSentenceId()
svgElement = GraphToSVG.writeSVG(svgTokens, svgEdges, self.outDir+"/svg/"+sentenceId+"-"+str(sentenceIndex)+"_learned.svg")
builder.svg("../svg/" + sentenceId + "-"+str(sentenceIndex)+"_learned.svg",svgElement.attrib["width"],svgElement.attrib["height"],id="learned_graph")
builder.lineBreak()
return stats
