Source Code for Module TEES.ExampleBuilders.ModifierExampleBuilder

  1  """ 
  2  Speculation and negation examples 
  3  """ 
  4  __version__ = "$Revision: 1.12 $" 
  5   
  6  import sys, os 
  7  import types 
  8  thisPath = os.path.dirname(os.path.abspath(__file__)) 
  9  sys.path.append(os.path.abspath(os.path.join(thisPath,".."))) 
 10  from ExampleBuilder import ExampleBuilder 
 11  import Utils.Libraries.PorterStemmer as PorterStemmer 
 12  from Core.IdSet import IdSet 
 13  import Core.ExampleUtils as ExampleUtils 
 14  #from Core.Gazetteer import Gazetteer 
 15  # For gold mapping 
 16  import Evaluators.EvaluateInteractionXML as EvaluateInteractionXML 
 17   
 18  # A list of speculation related words manually picked from the BioNLP'09 GE corpus 
 19  speculationWords = [ 
 20      'account', 'aim', 'almost', 'analysed', 'analyses', 'analysis', 'analyzed', 'appear',  
 21      'appeared', 'appears', 'argue', 'artifact', 'ascertain', 'asked', 'assayed',  
 22      'assessed', 'assumes', 'believed', 'can', 'candidates', 'clarify', 'clear',  
 23      'conclude', 'confirm', 'considered', 'could', 'define', 'delineate', 'determine',  
 24      'determined', 'elucidate', 'elucidating', 'establish', 'evaluate', 'evaluated',  
 25      'evaluates', 'evidence', 'examine', 'examined', 'explore', 'findings', 'hypothesis',  
 26      'hypothesize', 'hypothesized', 'idea', 'identification', 'implicated', 'implicates',  
 27      'implications', 'importance', 'important', 'indicate', 'indicated', 'indicators',  
 28      'information', 'insights', 'investigate', 'investigated', 'investigation', 'isolate',  
 29      'known', 'likely', 'may', 'measured', 'might', 'monitored', 'most', 'must',  
 30      'objective', 'obscure', 'observations', 'observed', 'partially', 'partly',  
 31      'performed', 'perhaps', 'play', 'plays', 'possible', 'postulated', 'potent',  
 32      'potential', 'potentially', 'probably', 'propose', 'proposed', 'putative',  
 33      'quantitated', 'reexamined', 'reported', 'revealed', 'role', 'screened', 'seemed',  
 34      'seems', 'shown', 'significantly', 'since', 'sought', 'studied', 'studies', 'study',  
 35      'suggest', 'suggested', 'suggesting', 'suggests', 'support', 'suspect', 'tested',  
 36      'thought', 'unclear', 'undefined', 'understand', 'unknown', 'whether'] 
 37   
 38 -def readWords(words): 
 39      if type(words) in types.StringTypes: 
 40          wordSet = set()     
 41          f = open(filename) 
 42          for line in f.readlines(): 
 43              wordSet.add(line.strip()) 
 44          f.close() 
 45      else: # assume it's a list 
 46          wordSet = set(words) 
 47      stemSet = set() 
 48      for word in wordSet: 
 49          stemSet.add(PorterStemmer.stem(word)) 
 50      return wordSet, stemSet 
 51   
 52  #def compareDependencyEdgesById(dep1, dep2): 
 53  #    """ 
 54  #    Dependency edges are sorted, so that the program behaves consistently 
 55  #    on the sama data between different runs. 
 56  #    """ 
 57  #    id1 = dep1[2].get("id") 
 58  #    id2 = dep2[2].get("id") 
 59  #    if id1 > id2: 
 60  #       return 1 
 61  #    elif id1 == id2: 
 62  #       return 0 
 63  #    else: # x<y 
 64  #       return -1 
 65   
 66 -class ModifierExampleBuilder(ExampleBuilder): 
 67 -    def __init__(self, style=None, classSet=None, featureSet=None, gazetteerFileName=None): 
 68          global speculationWords 
 69           
 70          if classSet == None: 
 71              classSet = IdSet(1) 
 72          assert( classSet.getId("neg") == 1 ) 
 73          if featureSet == None: 
 74              featureSet = IdSet() 
 75           
 76          self.specWords, self.specWordStems = readWords(speculationWords)  
 77           
 78          ExampleBuilder.__init__(self, classSet, featureSet) 
 79          #gazetteerFileName="/usr/share/biotext/GeniaChallenge/SharedTaskTriggerTest/gazetteer-train" 
 80          if gazetteerFileName!=None: 
 81              self.gazetteer=Gazetteer.loadGztr(gazetteerFileName) 
 82              print >> sys.stderr, "Loaded gazetteer from",gazetteerFileName 
 83          else: 
 84              self.gazetteer=None 
 85          self._setDefaultParameters({"classification":"multiclass", "speculation_words":True}, {"classification":("multiclass", "speculation", "negation")}) 
 86          self.styles = self.getParameters(style) 
 87       
 88 -    def getMergedEntityType(self, entities): 
 89          """ 
 90          If a single token belongs to multiple entities of different types, 
 91          a new, composite type is defined. This type is the alphabetically 
 92          ordered types of these entities joined with '---'. 
 93          """ 
 94          types = set() 
 95          for entity in entities: 
 96              types.add(entity.get("type")) 
 97          types = list(types) 
 98          types.sort() 
 99          typeString = "" 
100          for type in types: 
101              if typeString != "": 
102                  typeString += "---" 
103              typeString += type 
104          return typeString 
105       
106 -    def getTokenFeatures(self, token, sentenceGraph): 
107          """ 
108          Returns a list of features based on the attributes of a token. 
109          These can be used to define more complex features. 
110          """ 
111          # These features are cached when this method is first called 
112          # for a token. 
113          if self.tokenFeatures.has_key(token): 
114              return self.tokenFeatures[token] 
115          tokTxt=sentenceGraph.getTokenText(token) 
116          features = {} 
117          features["_txt_"+tokTxt]=1 
118          features["_POS_"+token.get("POS")]=1 
119          if self.styles["speculation_words"]:  
120              if tokTxt in self.specWords: 
121                  features["_spec"]=1 
122                  features["_spec_"+tokTxt]=1 
123              tokStem = PorterStemmer.stem(tokTxt) 
124              if tokStem in self.specWordStems: 
125                  features["_spec_stem"]=1 
126                  features["_spec_stem_"+tokStem]=1 
127          if sentenceGraph.tokenIsName[token]: 
128              features["_isName"]=1 
129              for entity in sentenceGraph.tokenIsEntityHead[token]: 
130                  if entity.get("isName") == "True": 
131                      features["_annType_"+entity.get("type")]=1 
132          if self.gazetteer and tokTxt.lower() in self.gazetteer: 
133              for label,weight in self.gazetteer[tokTxt.lower()].items(): 
134                  pass 
135                  #features["_knownLabel_"+label]=weight 
136          self.tokenFeatures[token] = features 
137          return features 
138       
139 -    def buildLinearOrderFeatures(self,sentenceGraph,index,tag,features): 
140          """ 
141          Linear features are built by marking token features with a tag 
142          that defines their relative position in the linear order. 
143          """ 
144          tag = "linear_"+tag 
145          for tokenFeature,w in self.getTokenFeatures(sentenceGraph.tokens[index], sentenceGraph).iteritems(): 
146              features[self.featureSet.getId(tag+tokenFeature)] = w 
147       
148 -    def buildExamplesFromGraph(self, sentenceGraph, outfile, goldGraph=None): 
149          """ 
150          Build one example for each token of the sentence 
151          """ 
152          examples = [] 
153          exampleIndex = 0 
154           
155          self.tokenFeatures = {} 
156   
157          if goldGraph != None: 
158              entityToGold = EvaluateInteractionXML.mapEntities(sentenceGraph.entities, goldGraph.entities) 
159           
160          namedEntityCount = 0 
161          entityCount = 0 
162          for entity in sentenceGraph.entities: 
163              if entity.get("isName") == "True": # known data which can be used for features 
164                  namedEntityCount += 1 
165              else: # known data which can be used for features 
166                  entityCount += 1 
167          namedEntityCountFeature = "nameCount_" + str(namedEntityCount) 
168          entityCountFeature = "entityCount_" + str(entityCount) 
169           
170          bagOfWords = {} 
171          for token in sentenceGraph.tokens: 
172              text = "bow_" + token.get("text") 
173              if not bagOfWords.has_key(text): 
174                  bagOfWords[text] = 0 
175              bagOfWords[text] += 1 
176              if sentenceGraph.tokenIsName[token]: 
177                  text = "ne_" + text 
178                  if not bagOfWords.has_key(text): 
179                      bagOfWords[text] = 0 
180                  bagOfWords[text] += 1 
181              if len(sentenceGraph.tokenIsEntityHead) > 0: 
182                  text = "ge_" + text 
183                  if not bagOfWords.has_key(text): 
184                      bagOfWords[text] = 0 
185                  bagOfWords[text] += 1 
186               
187              text = token.get("text") 
188              if self.styles["speculation_words"] and text in self.specWords: 
189                  if not bagOfWords.has_key("spec_bow_"+text): 
190                      bagOfWords["spec_bow_"+text] = 0 
191                  bagOfWords["spec_bow_"+text] += 1 
192                  bagOfWords["spec_sentence"] = 1 
193           
194          bowFeatures = {} 
195          for k,v in bagOfWords.iteritems(): 
196              bowFeatures[self.featureSet.getId(k)] = v 
197           
198          self.inEdgesByToken = {} 
199          self.outEdgesByToken = {} 
200          self.edgeSetByToken = {} 
201          for token in sentenceGraph.tokens: 
202              inEdges = sentenceGraph.dependencyGraph.getInEdges(token) 
203              self.inEdgesByToken[token] = inEdges 
204              outEdges = sentenceGraph.dependencyGraph.getOutEdges(token) 
205              self.outEdgesByToken[token] = outEdges 
206              self.edgeSetByToken[token] = set(inEdges + outEdges) 
207           
208          for entity in sentenceGraph.entities: 
209              #token = sentenceGraph.tokens[i] 
210              token = sentenceGraph.entityHeadTokenByEntity[entity] 
211              # Recognize only non-named entities (i.e. interaction words) 
212              if entity.get("isName") == "True": 
213                  continue 
214               
215              # CLASS 
216              if self.styles["classification"] == "multiclass": 
217                  task3Type = "multiclass" 
218                  categoryName = "" 
219                  if entity.get("negation") == "True": 
220                      categoryName += "negation" 
221                  if entity.get("speculation") == "True": 
222                      if categoryName != "": 
223                          categoryName += "---" 
224                      categoryName += "speculation" 
225                  if categoryName == "": 
226                      categoryName = "neg" 
227                  category = self.classSet.getId(categoryName)   
228              elif self.styles["classification"] == "speculation": 
229                  task3Type = "speculation" 
230                  if entity.get("speculation") == "True": 
231                      category = self.classSet.getId("speculation") 
232                  else: 
233                      category = 1 
234                  if goldGraph != None: 
235                      if len(entityToGold[entity]) > 0 and entityToGold[entity][0].get("speculation") == "True": 
236                          category = self.classSet.getId("speculation") 
237                      else: 
238                          category = 1 
239                  categoryName = self.classSet.getName(category) 
240              elif self.styles["classification"] == "negation": 
241                  task3Type = "negation" 
242                  if entity.get("negation") == "True": 
243                      category = self.classSet.getId("negation") 
244                  else: 
245                      category = 1 
246                  if goldGraph != None: 
247                      if len(entityToGold[entity]) > 0 and entityToGold[entity][0].get("negation") == "True": 
248                          category = self.classSet.getId("negation") 
249                      else: 
250                          category = 1 
251                  categoryName = self.classSet.getName(category) 
252              self.exampleStats.beginExample(categoryName) 
253   
254              # FEATURES 
255              features = {} 
256   
257              # ENTITY TYPE 
258              #entityType = self.classSet.getId(self.getMergedEntityType(entity)) 
259              #del self.classSet.Ids[self.getMergedEntityType(entity)] 
260  #IF LOCAL 
261              # There's a mistake here. The entityType should be the string, not 
262              # the id of the type. But there's also another issue. getMergedEntityType 
263              # expects a list, not an item. Therefore the type is always empty -> 
264              # types don't get used in classification. But this is the code used in 
265              # the publication, so it will now be published as is, and fixed in a later 
266              # release. 
267              # 
268              # Besides, using the classSet here generates an unneeded 
269              # additional class, that shows up in evaluations etc. However, to be  
270              # able to publish the exact models used for the publication experiments, 
271              # this can't be fixed so it breaks feature id consistency. Therefore I'll 
272              # now just remove the redundant class id from the classSet. 
273  #ENDIF             
274              #features[self.featureSet.getId(entityType)] = 1 
275               
276              features[self.featureSet.getId(namedEntityCountFeature)] = 1 
277              features[self.featureSet.getId(entityCountFeature)] = 1 
278              #for k,v in bagOfWords.iteritems(): 
279              #    features[self.featureSet.getId(k)] = v 
280              # pre-calculate bow _features_ 
281              features.update(bowFeatures) 
282               
283  #            for j in range(len(sentenceGraph.tokens)): 
284  #                text = "bow_" + sentenceGraph.tokens[j].get("text") 
285  #                if j < i: 
286  #                    features[self.featureSet.getId("bf_" + text)] = 1 
287  #                elif j > i: 
288  #                    features[self.featureSet.getId("af_" + text)] = 1 
289           
290              # Main features 
291              text = token.get("text") 
292              features[self.featureSet.getId("txt_"+text)] = 1 
293              features[self.featureSet.getId("POS_"+token.get("POS"))] = 1 
294              stem = PorterStemmer.stem(text) 
295              features[self.featureSet.getId("stem_"+stem)] = 1 
296              features[self.featureSet.getId("nonstem_"+text[len(stem):])] = 1 
297               
298              if self.styles["speculation_words"]: 
299                  if text in self.specWords: 
300                      features[self.featureSet.getId("ent_spec")] = 1 
301                  if stem in self.specWordStems: 
302                      features[self.featureSet.getId("ent_spec_stem")] = 1 
303               
304              # Linear order features 
305              for i in range(len(sentenceGraph.tokens)): 
306                  if token == sentenceGraph.tokens[i]: 
307                      break 
308              for index in [-3,-2,-1,1,2,3]: 
309                  if i + index > 0 and i + index < len(sentenceGraph.tokens): 
310                      self.buildLinearOrderFeatures(sentenceGraph, i + index, str(index), features) 
311               
312              # Content 
313              if i > 0 and text[0].isalpha() and text[0].isupper(): 
314                  features[self.featureSet.getId("upper_case_start")] = 1 
315              for j in range(len(text)): 
316                  if j > 0 and text[j].isalpha() and text[j].isupper(): 
317                      features[self.featureSet.getId("upper_case_middle")] = 1 
318                  # numbers and special characters 
319                  if text[j].isdigit(): 
320                      features[self.featureSet.getId("has_digits")] = 1 
321                      if j > 0 and text[j-1] == "-": 
322                          features[self.featureSet.getId("has_hyphenated_digit")] = 1 
323                  elif text[j] == "-": 
324                      features[self.featureSet.getId("has_hyphen")] = 1 
325                  elif text[j] == "/": 
326                      features[self.featureSet.getId("has_fslash")] = 1 
327                  elif text[j] == "\\": 
328                      features[self.featureSet.getId("has_bslash")] = 1 
329                  # duplets 
330                  if j > 0: 
331                      features[self.featureSet.getId("dt_"+text[j-1:j+1].lower())] = 1 
332                  # triplets 
333                  if j > 1: 
334                      features[self.featureSet.getId("tt_"+text[j-2:j+1].lower())] = 1 
335               
336              # Attached edges (Hanging in and out edges) 
337              t1InEdges = self.inEdgesByToken[token] 
338              for edge in t1InEdges: 
339                  edgeType = edge[2].get("type") 
340                  features[self.featureSet.getId("t1HIn_"+edgeType)] = 1 
341                  features[self.featureSet.getId("t1HIn_"+edge[0].get("POS"))] = 1 
342                  features[self.featureSet.getId("t1HIn_"+edgeType+"_"+edge[0].get("POS"))] = 1 
343                  tokenText = sentenceGraph.getTokenText(edge[0]) 
344                  features[self.featureSet.getId("t1HIn_"+tokenText)] = 1 
345                  features[self.featureSet.getId("t1HIn_"+edgeType+"_"+tokenText)] = 1 
346              t1OutEdges = self.outEdgesByToken[token] 
347              for edge in t1OutEdges: 
348                  edgeType = edge[2].get("type") 
349                  features[self.featureSet.getId("t1HOut_"+edgeType)] = 1 
350                  features[self.featureSet.getId("t1HOut_"+edge[1].get("POS"))] = 1 
351                  features[self.featureSet.getId("t1HOut_"+edgeType+"_"+edge[1].get("POS"))] = 1 
352                  tokenText = sentenceGraph.getTokenText(edge[1]) 
353                  features[self.featureSet.getId("t1HOut_"+tokenText)] = 1 
354                  features[self.featureSet.getId("t1HOut_"+edgeType+"_"+tokenText)] = 1 
355   
356              self.buildChains(token, sentenceGraph, features) 
357                
358              extra = {"xtype":"task3","t3type":task3Type,"t":token.get("id"),"entity":entity.get("id")} 
359              #examples.append( (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) ) 
360              example = (sentenceGraph.getSentenceId()+".x"+str(exampleIndex),category,features,extra) 
361              ExampleUtils.appendExamples([example], outfile) 
362              exampleIndex += 1             
363              self.exampleStats.endExample() 
364          #return examples 
365          return exampleIndex 
366       
367 -    def buildChains(self,token,sentenceGraph,features,depthLeft=3,chain="",visited=None): 
368          if depthLeft == 0: 
369              return 
370          strDepthLeft = "dist_" + str(depthLeft) 
371           
372          if visited == None: 
373              visited = set() 
374   
375          inEdges = self.inEdgesByToken[token] 
376          outEdges = self.outEdgesByToken[token] 
377          edgeSet = visited.union(self.edgeSetByToken[token]) 
378          for edge in inEdges: 
379              if not edge in visited: 
380                  edgeType = edge[2].get("type") 
381                  features[self.featureSet.getId("dep_"+strDepthLeft+edgeType)] = 1 
382   
383                  nextToken = edge[0] 
384                  for tokenFeature,w in self.getTokenFeatures(nextToken, sentenceGraph).iteritems(): 
385                      features[self.featureSet.getId(strDepthLeft + tokenFeature)] = w 
386                   
387                  features[self.featureSet.getId("chain_dist_"+strDepthLeft+chain+"-frw_"+edgeType)] = 1 
388                  self.buildChains(nextToken,sentenceGraph,features,depthLeft-1,chain+"-frw_"+edgeType,edgeSet) 
389   
390          for edge in outEdges: 
391              if not edge in visited: 
392                  edgeType = edge[2].get("type") 
393                  features[self.featureSet.getId("dep_dist_"+strDepthLeft+edgeType)] = 1 
394   
395                  nextToken = edge[1] 
396                  for tokenFeature,w in self.getTokenFeatures(nextToken, sentenceGraph).iteritems(): 
397                      features[self.featureSet.getId(strDepthLeft + tokenFeature)] = w 
398                   
399                  features[self.featureSet.getId("chain_dist_"+strDepthLeft+chain+"-rev_"+edgeType)] = 1 
400                  self.buildChains(nextToken,sentenceGraph,features,depthLeft-1,chain+"-rev_"+edgeType,edgeSet) 
401