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
features["_POS_"+token.get("POS")]=1
if self.styles["speculation_words"]:
if tokTxt in self.specWords:
features["_spec"]=1
features["_spec_"+tokTxt]=1
tokStem = PorterStemmer.stem(tokTxt)
if tokStem in self.specWordStems:
features["_spec_stem"]=1
features["_spec_stem_"+tokStem]=1
if sentenceGraph.tokenIsName[token]:
features["_given"]=1
for entity in sentenceGraph.tokenIsEntityHead[token]:
if entity.get("given") == "True":
features["_annType_"+entity.get("type")]=1
if self.gazetteer and tokTxt.lower() in self.gazetteer:
for label,weight in self.gazetteer[tokTxt.lower()].items():
pass
#features["_knownLabel_"+label]=weight
self.tokenFeatures[token] = features
return features
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
features["_POS_" + token.get("POS")] = 1
if self.styles["speculation_words"]:
if tokTxt in self.specWords:
features["_spec"] = 1
features["_spec_" + tokTxt] = 1
tokStem = PorterStemmer.stem(tokTxt)
if tokStem in self.specWordStems:
features["_spec_stem"] = 1
features["_spec_stem_" + tokStem] = 1
if sentenceGraph.tokenIsName[token]:
features["_given"] = 1
for entity in sentenceGraph.tokenIsEntityHead[token]:
if entity.get("given") == "True":
features["_annType_" + entity.get("type")] = 1
if self.gazetteer and tokTxt.lower() in self.gazetteer:
for label, weight in self.gazetteer[tokTxt.lower()].items():
pass
#features["_knownLabel_"+label]=weight
self.tokenFeatures[token] = features
return features
def mapSplits(splits, string, stringOffset):
"""
Maps substrings to a string, and stems them
"""
begin = 0
tuples = []
for split in splits:
offset = string.find(split, begin)
assert offset != -1
tuples.append( (split, PorterStemmer.stem(split), (offset,len(split))) )
begin = offset + len(split)
return tuples
def getTokenFeatures(self, token, sentenceGraph, text=True, POS=True, annotatedType=True, stem=False, ontology=True):
"""
Token features are features describing an isolated word token. These subfeatures are often merged into
such features like n-grams. This method produces and caches a set of feature names for a token in
the sentenceGraph sentence. The various flags can be used to choose which attributes will be included in the
feature name list.
@type token: cElementTree.Element
@param token: a word token
@type sentenceGraph: SentenceGraph
@param sentenceGraph: the sentence to which the token belongs
@type text: boolean
@type POS: boolean
@type annotatedType: boolean
@type stem: boolean
@type ontology: boolean
"""
callId = token.get("id") + str(text) + str(POS) + str(annotatedType) + str(stem) + str(ontology)
if self.tokenFeatures.has_key(callId):
return self.tokenFeatures[callId]
featureList = []
if text:
featureList.append("txt_"+sentenceGraph.getTokenText(token))
if (not self.maskNamedEntities) and sentenceGraph.tokenIsName[token]:
featureList.append("txt_"+token.get("text"))
if POS:
pos = token.get("POS")
if pos.find("_") != None and self.maximum:
for split in pos.split("_"):
featureList.append("POS_"+split)
featureList.append("POS_"+pos)
#if self.getPOSSuperType(pos) != "":
# featureList.append("POSX_"+self.getPOSSuperType(pos))
if annotatedType and not self.noAnnType:
annTypes = self.getTokenAnnotatedType(token, sentenceGraph)
if "noAnnType" in annTypes and not self.maximum:
annTypes.remove("noAnnType")
for annType in annTypes:
featureList.append("annType_"+annType)
if ontology and (self.ontologyFeatureBuilder != None):
for annType in annTypes:
featureList.extend(self.ontologyFeatureBuilder.getParents(annType))
if stem:
featureList.append("stem_" + PorterStemmer.stem(sentenceGraph.getTokenText(token)))
if self.style != None and self.style["metamap"]:
metamapFeatureDict = {}
self.getMetaMapFeatures(token, sentenceGraph, metamapFeatureDict)
featureList.extend(sorted(metamapFeatureDict.keys()))
self.tokenFeatures[callId] = featureList
return featureList
def mapSplits(splits, string, stringOffset):
"""
Maps substrings to a string, and stems them
"""
begin = 0
tuples = []
for split in splits:
offset = string.find(split, begin)
assert offset != -1
tuples.append((split, PorterStemmer.stem(split), (offset, len(split))))
begin = offset + len(split)
return tuples
def readWords(words):
if type(words) in types.StringTypes:
wordSet = set()
f = open(filename)
for line in f.readlines():
wordSet.add(line.strip())
f.close()
else: # assume it's a list
wordSet = set(words)
stemSet = set()
for word in wordSet:
stemSet.add(PorterStemmer.stem(word))
return wordSet, stemSet
def readWords(words):
if type(words) in types.StringTypes:
wordSet = set()
f = open(filename)
for line in f.readlines():
wordSet.add(line.strip())
f.close()
else: # assume it's a list
wordSet = set(words)
stemSet = set()
for word in wordSet:
stemSet.add(PorterStemmer.stem(word))
return wordSet, stemSet
def getTriggers(corpus):
"""
Returns a dictionary of "entity type"->"entity text"->"count"
"""
corpus = ETUtils.ETFromObj(corpus)
trigDict = {}
for entity in corpus.getroot().getiterator("entity"):
if entity.get("isName") == "True":
continue
eType = entity.get("type")
if not trigDict.has_key(eType):
trigDict[eType] = {}
eText = entity.get("text")
eText = PorterStemmer.stem(eText)
if not trigDict[eType].has_key(eText):
trigDict[eType][eText] = 0
trigDict[eType][eText] += 1
return trigDict
def getTriggers(corpus):
"""
Returns a dictionary of "entity type"->"entity text"->"count"
"""
corpus = ETUtils.ETFromObj(corpus)
trigDict = {}
for entity in corpus.getroot().getiterator("entity"):
if entity.get("given") == "True":
continue
eType = entity.get("type")
if not trigDict.has_key(eType):
trigDict[eType] = {}
eText = entity.get("text")
eText = PorterStemmer.stem(eText)
if not trigDict[eType].has_key(eText):
trigDict[eType][eText] = 0
trigDict[eType][eText] += 1
return trigDict
def buildFeatures(self, token, linear=True, chains=True):
sentenceGraph = self.sentenceGraph
tokenIndex = None
for i in range(len(self.sentenceGraph.tokens)):
if token == self.sentenceGraph.tokens[i]:
tokenIndex = i
break
assert tokenIndex != None
token = self.sentenceGraph.tokens[tokenIndex]
#if not "names" in self.styles:
self.setFeature(self.namedEntityCountFeature, 1)
#self.features.update(self.bowFeatures) # Note! these do not get tagged
# 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")
self.setFeature("txt_" + text, 1)
self.setFeature("POS_" + token.get("POS"), 1)
stem = PorterStemmer.stem(text)
self.setFeature("stem_" + stem, 1)
self.setFeature("nonstem_" + text[len(stem):], 1)
# Linear order features
if linear:
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))
# Content
if i > 0 and text[0].isalpha() and text[0].isupper():
self.setFeature("upper_case_start", 1)
for j in range(len(text)):
if j > 0 and text[j].isalpha() and text[j].isupper():
self.setFeature("upper_case_middle", 1)
# numbers and special characters
if text[j].isdigit():
self.setFeature("has_digits", 1)
if j > 0 and text[j - 1] == "-":
self.setFeature("has_hyphenated_digit", 1)
elif text[j] == "-":
self.setFeature("has_hyphen", 1)
elif text[j] == "/":
self.setFeature("has_fslash", 1)
elif text[j] == "\\":
self.setFeature("has_bslash", 1)
# duplets
if j > 0:
self.setFeature("dt_" + text[j - 1:j + 1].lower(), 1)
# triplets
if j > 1:
self.setFeature("tt_" + text[j - 2:j + 1].lower(), 1)
# chains
if chains:
self.buildChains(token, sentenceGraph)
def buildExamplesFromGraph(self, sentenceGraph, outfile, goldGraph=None, structureAnalyzer=None):
"""
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 0 #[]
#examples = []
exampleIndex = 0
self.tokenFeatures = {}
self.tokenFeatureWeights = {}
# determine (manually or automatically) the setting for whether sentences with no given entities should be skipped
buildForNameless = False
if structureAnalyzer and not structureAnalyzer.hasGroupClass("GIVEN", "ENTITY"): # no given entities points to no separate NER program being used
buildForNameless = True
if self.styles["build_for_nameless"]: # manually force the setting
buildForNameless = True
if self.styles["skip_for_nameless"]: # manually force the setting
buildForNameless = False
# determine whether sentences with no given entities should be skipped
namedEntityHeadTokens = []
if not self.styles["names"]:
namedEntityCount = 0
for entity in sentenceGraph.entities:
if entity.get("given") == "True": # known data which can be used for features
namedEntityCount += 1
namedEntityCountFeature = "nameCount_" + str(namedEntityCount)
# NOTE!!! This will change the number of examples and omit
# all triggers (positive and negative) from sentences which
# have no NE:s, possibly giving a too-optimistic performance
# value. Such sentences can still have triggers from intersentence
# interactions, but as such events cannot be recovered anyway,
# looking for these triggers would be pointless.
if namedEntityCount == 0 and not buildForNameless: # no names, no need for triggers
return 0 #[]
if self.styles["pos_pairs"]:
namedEntityHeadTokens = self.getNamedEntityHeadTokens(sentenceGraph)
else:
for key in sentenceGraph.tokenIsName.keys():
sentenceGraph.tokenIsName[key] = False
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 in sorted(bagOfWords.keys()):
bowFeatures[self.featureSet.getId(k)] = bagOfWords[k]
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 = sentenceGraph.dependencyGraph.getInEdges(token)
#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 = sentenceGraph.dependencyGraph.getOutEdges(token)
#outEdges.sort(compareDependencyEdgesById)
self.outEdgesByToken[token] = outEdges
self.edgeSetByToken[token] = set(inEdges + outEdges)
for i in range(len(sentenceGraph.tokens)):
token = sentenceGraph.tokens[i]
# CLASS
if len(sentenceGraph.tokenIsEntityHead[token]) > 0:
categoryName, entityIds = self.getMergedEntityType(sentenceGraph.tokenIsEntityHead[token])
else:
categoryName, entityIds = "neg", None
self.exampleStats.beginExample(categoryName)
# Recognize only non-named entities (i.e. interaction words)
if sentenceGraph.tokenIsName[token] and not self.styles["names"] and not self.styles["all_tokens"]:
self.exampleStats.filter("name")
self.exampleStats.endExample()
continue
# if "selftrain_limits" in self.styles:
# # any predicted entity not part of the self-training set causes example to be rejected
# filtered = False
# for entity in sentenceGraph.tokenIsEntityHead[token]:
# if entity.get("selftrain") == "False":
# self.exampleStats.filter("selftrain_limits")
# self.exampleStats.endExample()
# filtered = True
# break
# if filtered:
# continue
# if "selftrain_group" in self.styles:
# # any predicted entity not part of the self-training set causes example to be rejected
# filtered = False
# for entity in sentenceGraph.tokenIsEntityHead[token]:
# if entity.get("selftraingroup") not in self.selfTrainGroups:
# self.exampleStats.filter("selftrain_group")
# self.exampleStats.endExample()
# filtered = True
# break
# if filtered:
# continue
if self.styles["pos_only"] and categoryName == "neg":
self.exampleStats.filter("pos_only")
self.exampleStats.endExample()
continue
category = self.classSet.getId(categoryName)
if category == None:
self.exampleStats.filter("undefined_class")
self.exampleStats.endExample()
continue
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"}
# if entityIds != None:
# extra["goldIds"] = entityIds
# #examples.append( (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) )
# ExampleUtils.appendExamples([(sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra)], outfile)
# exampleIndex += 1
# continue
# FEATURES
features = {}
if not self.styles["names"]:
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)
# 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
## Subspan features
#textLower = text.lower()
#for i in range(1, len(textLower)):
# features[self.featureSet.getId("subspanbegin"+str(i)+"_"+textLower[0:i])] = 1
# features[self.featureSet.getId("subspanend"+str(i)+"_"+textLower[-i:])] = 1
# Substring features
for string in text.split("-"):
stringLower = string.lower()
features[self.featureSet.getId("substring_"+stringLower)] = 1
features[self.featureSet.getId("substringstem_"+PorterStemmer.stem(stringLower))] = 1
# Linear order features
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)
# Linear n-grams
if self.styles["linear_ngrams"]:
self.buildLinearNGram(max(0, i-1), i, sentenceGraph, features)
self.buildLinearNGram(max(0, i-2), i, sentenceGraph, features)
if self.styles["phospho"]:
if text.find("hospho") != -1:
features[self.featureSet.getId("phospho_found")] = 1
features[self.featureSet.getId("begin_"+text[0:2].lower())] = 1
features[self.featureSet.getId("begin_"+text[0:3].lower())] = 1
if self.styles["bb_features"]:
if text.lower() in self.bacteriaTokens:
features[self.featureSet.getId("lpsnBacToken")] = 1
# 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
# quadruplets (don't work, slight decrease (0.5 pp) on f-score
#if j > 2:
# features[self.featureSet.getId("qt_"+text[j-3: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
tokenStem = PorterStemmer.stem(tokenText)
features[self.featureSet.getId("t1HIn_"+tokenStem)] = 1
features[self.featureSet.getId("t1HIn_"+edgeType+"_"+tokenStem)] = 1
features[self.featureSet.getId("t1HIn_"+norStem+"_"+edgeType+"_"+tokenStem)] = 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
tokenStem = PorterStemmer.stem(tokenText)
features[self.featureSet.getId("t1HOut_"+tokenStem)] = 1
features[self.featureSet.getId("t1HOut_"+edgeType+"_"+tokenStem)] = 1
features[self.featureSet.getId("t1HOut_"+norStem+"_"+edgeType+"_"+tokenStem)] = 1
# REL features
if self.styles["rel_features"]:
self.relFeatureBuilder.setFeatureVector(features)
self.relFeatureBuilder.buildAllFeatures(sentenceGraph.tokens, i)
self.relFeatureBuilder.setFeatureVector(None)
# DDI13 features
if self.styles["ddi13_features"]:
for index in range(len(normalizedText)):
features[self.featureSet.getId("ddi13_fromstart" + str(index) + "_" + normalizedText[:index+1])] = 1
features[self.featureSet.getId("ddi13_fromend" + str(index) + "_" + normalizedText[index:])] = 1
if self.styles["drugbank_features"]:
self.drugFeatureBuilder.setFeatureVector(features)
self.drugFeatureBuilder.tag = "ddi_"
self.drugFeatureBuilder.buildDrugFeatures(token)
self.drugFeatureBuilder.setFeatureVector(None)
#self.wordNetFeatureBuilder.getTokenFeatures("show", "VBP")
#tokTxt = token.get("text")
#tokPOS = token.get("POS")
#wordNetFeatures = []
#wordNetFeatures = self.wordNetFeatureBuilder.getTokenFeatures(tokTxt, tokPOS)
#self.wordNetFeatureBuilder.getTokenFeatures(tokTxt, tokPOS)
if self.styles["wordnet"]:
tokTxt = token.get("text")
tokPOS = token.get("POS")
wordNetFeatures = self.wordNetFeatureBuilder.getTokenFeatures(tokTxt, tokPOS)
for wordNetFeature in wordNetFeatures:
#print wordNetFeature,
features[self.featureSet.getId("WN_"+wordNetFeature)] = 1
#print
if self.styles["giuliano"]:
self.giulianoFeatureBuilder.setFeatureVector(features)
self.giulianoFeatureBuilder.buildTriggerFeatures(token, sentenceGraph)
self.giulianoFeatureBuilder.setFeatureVector(None)
extra = {"xtype":"token","t":token.get("id")}
if self.styles["bb_features"]:
extra["trigex"] = "bb" # Request trigger extension in ExampleWriter
if self.styles["epi_merge_negated"]:
extra["unmergeneg"] = "epi" # Request trigger type unmerging
if entityIds != None:
extra["goldIds"] = entityIds # The entities to which this example corresponds
#examples.append( (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) )
# chains
self.buildChains(token, sentenceGraph, features)
if self.styles["pos_pairs"]:
self.buildPOSPairs(token, namedEntityHeadTokens, features)
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,
structureAnalyzer=None):
"""
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 0 #[]
#examples = []
exampleIndex = 0
self.tokenFeatures = {}
self.tokenFeatureWeights = {}
# determine (manually or automatically) the setting for whether sentences with no given entities should be skipped
buildForNameless = False
if structureAnalyzer and not structureAnalyzer.hasGroupClass(
"GIVEN", "ENTITY"
): # no given entities points to no separate NER program being used
buildForNameless = True
if self.styles["build_for_nameless"]: # manually force the setting
buildForNameless = True
if self.styles["skip_for_nameless"]: # manually force the setting
buildForNameless = False
# determine whether sentences with no given entities should be skipped
namedEntityHeadTokens = []
if not self.styles["names"]:
namedEntityCount = 0
for entity in sentenceGraph.entities:
if entity.get(
"given"
) == "True": # known data which can be used for features
namedEntityCount += 1
namedEntityCountFeature = "nameCount_" + str(namedEntityCount)
# NOTE!!! This will change the number of examples and omit
# all triggers (positive and negative) from sentences which
# have no NE:s, possibly giving a too-optimistic performance
# value. Such sentences can still have triggers from intersentence
# interactions, but as such events cannot be recovered anyway,
# looking for these triggers would be pointless.
if namedEntityCount == 0 and not buildForNameless: # no names, no need for triggers
return 0 #[]
if self.styles["pos_pairs"]:
namedEntityHeadTokens = self.getNamedEntityHeadTokens(
sentenceGraph)
else:
for key in sentenceGraph.tokenIsName.keys():
sentenceGraph.tokenIsName[key] = False
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 in sorted(bagOfWords.keys()):
bowFeatures[self.featureSet.getId(k)] = bagOfWords[k]
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 = sentenceGraph.dependencyGraph.getInEdges(token)
#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 = sentenceGraph.dependencyGraph.getOutEdges(token)
#outEdges.sort(compareDependencyEdgesById)
self.outEdgesByToken[token] = outEdges
self.edgeSetByToken[token] = set(inEdges + outEdges)
for i in range(len(sentenceGraph.tokens)):
token = sentenceGraph.tokens[i]
# CLASS
if len(sentenceGraph.tokenIsEntityHead[token]) > 0:
categoryName, entityIds = self.getMergedEntityType(
sentenceGraph.tokenIsEntityHead[token])
else:
categoryName, entityIds = "neg", None
self.exampleStats.beginExample(categoryName)
# Recognize only non-named entities (i.e. interaction words)
if sentenceGraph.tokenIsName[token] and not self.styles[
"names"] and not self.styles["all_tokens"]:
self.exampleStats.filter("name")
self.exampleStats.endExample()
continue
# if "selftrain_limits" in self.styles:
# # any predicted entity not part of the self-training set causes example to be rejected
# filtered = False
# for entity in sentenceGraph.tokenIsEntityHead[token]:
# if entity.get("selftrain") == "False":
# self.exampleStats.filter("selftrain_limits")
# self.exampleStats.endExample()
# filtered = True
# break
# if filtered:
# continue
# if "selftrain_group" in self.styles:
# # any predicted entity not part of the self-training set causes example to be rejected
# filtered = False
# for entity in sentenceGraph.tokenIsEntityHead[token]:
# if entity.get("selftraingroup") not in self.selfTrainGroups:
# self.exampleStats.filter("selftrain_group")
# self.exampleStats.endExample()
# filtered = True
# break
# if filtered:
# continue
if self.styles["pos_only"] and categoryName == "neg":
self.exampleStats.filter("pos_only")
self.exampleStats.endExample()
continue
category = self.classSet.getId(categoryName)
if category == None:
self.exampleStats.filter("undefined_class")
self.exampleStats.endExample()
continue
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"}
# if entityIds != None:
# extra["goldIds"] = entityIds
# #examples.append( (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) )
# ExampleUtils.appendExamples([(sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra)], outfile)
# exampleIndex += 1
# continue
# FEATURES
features = {}
if not self.styles["names"]:
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)
# 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
## Subspan features
#textLower = text.lower()
#for i in range(1, len(textLower)):
# features[self.featureSet.getId("subspanbegin"+str(i)+"_"+textLower[0:i])] = 1
# features[self.featureSet.getId("subspanend"+str(i)+"_"+textLower[-i:])] = 1
# Substring features
for string in text.split("-"):
stringLower = string.lower()
features[self.featureSet.getId("substring_" + stringLower)] = 1
features[self.featureSet.getId(
"substringstem_" + PorterStemmer.stem(stringLower))] = 1
if not self.styles["no_context"]:
# Linear order features
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)
# Linear n-grams
if self.styles["linear_ngrams"]:
self.buildLinearNGram(max(0, i - 1), i, sentenceGraph,
features)
self.buildLinearNGram(max(0, i - 2), i, sentenceGraph,
features)
if self.styles["phospho"]:
if text.find("hospho") != -1:
features[self.featureSet.getId("phospho_found")] = 1
features[self.featureSet.getId("begin_" +
text[0:2].lower())] = 1
features[self.featureSet.getId("begin_" +
text[0:3].lower())] = 1
if self.styles["bb_features"]:
if text.lower() in self.bacteriaTokens:
features[self.featureSet.getId("lpsnBacToken")] = 1
# 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
# quadruplets (don't work, slight decrease (0.5 pp) on f-score
#if j > 2:
# features[self.featureSet.getId("qt_"+text[j-3:j+1].lower())] = 1
# Attached edges (Hanging in and out edges)
if not self.styles["no_context"]:
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
tokenStem = PorterStemmer.stem(tokenText)
features[self.featureSet.getId("t1HIn_" + tokenStem)] = 1
features[self.featureSet.getId("t1HIn_" + edgeType + "_" +
tokenStem)] = 1
features[self.featureSet.getId("t1HIn_" + norStem + "_" +
edgeType + "_" +
tokenStem)] = 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
tokenStem = PorterStemmer.stem(tokenText)
features[self.featureSet.getId("t1HOut_" + tokenStem)] = 1
features[self.featureSet.getId("t1HOut_" + edgeType + "_" +
tokenStem)] = 1
features[self.featureSet.getId("t1HOut_" + norStem + "_" +
edgeType + "_" +
tokenStem)] = 1
# REL features
if self.styles["rel_features"]:
self.relFeatureBuilder.setFeatureVector(features)
self.relFeatureBuilder.buildAllFeatures(
sentenceGraph.tokens, i)
self.relFeatureBuilder.setFeatureVector(None)
# DDI13 features
if self.styles["ddi13_features"]:
for index in range(len(normalizedText)):
features[self.featureSet.getId("ddi13_fromstart" +
str(index) + "_" +
normalizedText[:index +
1])] = 1
features[self.featureSet.getId("ddi13_fromend" +
str(index) + "_" +
normalizedText[index:])] = 1
if self.styles["drugbank_features"]:
self.drugFeatureBuilder.setFeatureVector(features)
self.drugFeatureBuilder.tag = "ddi_"
self.drugFeatureBuilder.buildDrugFeatures(token)
self.drugFeatureBuilder.setFeatureVector(None)
#self.wordNetFeatureBuilder.getTokenFeatures("show", "VBP")
#tokTxt = token.get("text")
#tokPOS = token.get("POS")
#wordNetFeatures = []
#wordNetFeatures = self.wordNetFeatureBuilder.getTokenFeatures(tokTxt, tokPOS)
#self.wordNetFeatureBuilder.getTokenFeatures(tokTxt, tokPOS)
if self.styles["wordnet"]:
tokTxt = token.get("text")
tokPOS = token.get("POS")
wordNetFeatures = self.wordNetFeatureBuilder.getTokenFeatures(
tokTxt, tokPOS)
for wordNetFeature in wordNetFeatures:
#print wordNetFeature,
features[self.featureSet.getId("WN_" + wordNetFeature)] = 1
#print
if self.styles["giuliano"]:
self.giulianoFeatureBuilder.setFeatureVector(features)
self.giulianoFeatureBuilder.buildTriggerFeatures(
token, sentenceGraph)
self.giulianoFeatureBuilder.setFeatureVector(None)
if self.styles["ontobiotope_features"]:
self.ontobiotopeFeatureBuilder.setFeatureVector(features)
self.ontobiotopeFeatureBuilder.buildOBOFeaturesForToken(token)
self.ontobiotopeFeatureBuilder.setFeatureVector(None)
extra = {"xtype": "token", "t": token.get("id")}
if self.styles["bb_features"]:
extra[
"trigex"] = "bb" # Request trigger extension in ExampleWriter
if self.styles["epi_merge_negated"]:
extra["unmergeneg"] = "epi" # Request trigger type unmerging
if entityIds != None:
extra[
"goldIds"] = entityIds # The entities to which this example corresponds
#examples.append( (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) )
if self.styles["bb_spans"]:
for span in sentenceGraph.sentenceElement.iter("span"):
if span.get("headOffset") != token.get("charOffset"):
continue
#if span.get("source") != "spec":
# continue
#print span.get("headOffset"), token.get("charOffset"), span.get("source"), token.get("id")
features[self.featureSet.getId("span_found")] = 1
features[self.featureSet.getId(
"span_count")] = 1 + features.get(
self.featureSet.getId("span_count"), 0)
features[self.featureSet.getId("span_identifier" +
span.get("identifier"))] = 1
features[self.featureSet.getId("span_type" +
span.get("type"))] = 1
features[self.featureSet.getId("span_category" +
span.get("category"))] = 1
features[self.featureSet.getId("span_source" +
span.get("source"))] = 1
if "define_offset" in extra:
prevOffset = [
int(x) for x in extra["define_offset"].split("-")
]
assert len(prevOffset) == 2
newOffset = [
int(x) for x in span.get("charOffset").split("-")
]
assert len(newOffset) == 2
prevOffsetRange = abs(prevOffset[0] - prevOffset[1])
newOffsetRange = abs(newOffset[0] - newOffset[1])
if newOffsetRange > prevOffsetRange:
extra["define_offset"] = span.get("charOffset")
else:
extra["define_offset"] = span.get("charOffset")
features[self.featureSet.getId("span_count_" + str(
features.get(self.featureSet.getId("span_count"), 0)))] = 1
# chains
if not self.styles["no_context"]:
self.buildChains(token, sentenceGraph, features)
if self.styles["pos_pairs"]:
self.buildPOSPairs(token, namedEntityHeadTokens, features)
if self.styles["wordvector"]:
self.wordVectorFeatureBuilder.setFeatureVector(features)
self.wordVectorFeatureBuilder.buildFeatures(token)
self.wordVectorFeatureBuilder.setFeatureVector(None)
example = (sentenceGraph.getSentenceId() + ".x" +
str(exampleIndex), category, features, extra)
ExampleUtils.appendExamples([example], outfile)
exampleIndex += 1
self.exampleStats.endExample()
#return examples
return exampleIndex
def stemTokens(self):
for token in self.tokensById.values():
token.stem = stemmer.stem(token.text)
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, 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 buildFeatures(self, token, linear=True, chains=True):
sentenceGraph = self.sentenceGraph
tokenIndex = None
for i in range(len(self.sentenceGraph.tokens)):
if token == self.sentenceGraph.tokens[i]:
tokenIndex = i
break
assert tokenIndex != None
token = self.sentenceGraph.tokens[tokenIndex]
# if not "names" in self.styles:
self.setFeature(self.namedEntityCountFeature, 1)
# self.features.update(self.bowFeatures) # Note! these do not get tagged
# 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")
self.setFeature("txt_" + text, 1)
self.setFeature("POS_" + token.get("POS"), 1)
stem = PorterStemmer.stem(text)
self.setFeature("stem_" + stem, 1)
self.setFeature("nonstem_" + text[len(stem) :], 1)
# Linear order features
if linear:
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))
# Content
if i > 0 and text[0].isalpha() and text[0].isupper():
self.setFeature("upper_case_start", 1)
for j in range(len(text)):
if j > 0 and text[j].isalpha() and text[j].isupper():
self.setFeature("upper_case_middle", 1)
# numbers and special characters
if text[j].isdigit():
self.setFeature("has_digits", 1)
if j > 0 and text[j - 1] == "-":
self.setFeature("has_hyphenated_digit", 1)
elif text[j] == "-":
self.setFeature("has_hyphen", 1)
elif text[j] == "/":
self.setFeature("has_fslash", 1)
elif text[j] == "\\":
self.setFeature("has_bslash", 1)
# duplets
if j > 0:
self.setFeature("dt_" + text[j - 1 : j + 1].lower(), 1)
# triplets
if j > 1:
self.setFeature("tt_" + text[j - 2 : j + 1].lower(), 1)
# chains
if chains:
self.buildChains(token, sentenceGraph)
def stemTokens(self):
for token in self.tokensById.values():
token.stem = stemmer.stem(token.text)
