}
Random rand = new Random();
//TRY SMOTE algorithm here to generate more minority class samples
SMOTEplus smote = new SMOTEplus((int)Math.ceil(this.smoteNumOfNeighbors));
// classify tokens within each sentence
for (Sentence sentence : JCasUtil.selectCovered(jCas, Sentence.class, segment)) {
List<BaseToken> tokens = JCasUtil.selectCovered(jCas, BaseToken.class, sentence);
// during training, the list of all outcomes for the tokens
List<String> outcomes;
if (this.isTraining()) {
List<EventMention> events = Lists.newArrayList();
for (EventMention event : JCasUtil.selectCovered(jCas, EventMention.class, sentence)) {
if (event.getClass().equals(EventMention.class)) {
events.add(event);
}
}
outcomes = this.eventChunking.createOutcomes(jCas, tokens, events);
}
// during prediction, the list of outcomes predicted so far
else {
outcomes = new ArrayList<String>();
}
// get BIO entity tags for each entity type
int[] entityTypeIDs = new int[] {
CONST.NE_TYPE_ID_ANATOMICAL_SITE,
CONST.NE_TYPE_ID_DISORDER,
CONST.NE_TYPE_ID_DRUG,
CONST.NE_TYPE_ID_FINDING,
CONST.NE_TYPE_ID_PROCEDURE,
CONST.NE_TYPE_ID_UNKNOWN };
List<IdentifiedAnnotation> entities;
if (this.isTraining()) {
entities = Lists.newArrayList();
for (IdentifiedAnnotation entity : JCasUtil.selectCovered(jCas, IdentifiedAnnotation.class, sentence)) {
if (!entity.getClass().equals(EventMention.class)) {
entities.add(entity);
}
}
} else {
entities = JCasUtil.selectCovered(jCas, IdentifiedAnnotation.class, sentence);
}
List<ChunkingExtractor> chunkingExtractors = Lists.newArrayList();
for (int typeID : entityTypeIDs) {
Predicate<IdentifiedAnnotation> hasTypeID = hasEntityType(typeID);
List<IdentifiedAnnotation> subEntities = Lists.newArrayList(Iterables.filter(entities, hasTypeID));
chunkingExtractors.add(new ChunkingExtractor("EntityTag", this.entityChunking, jCas, tokens, subEntities));
}
// add extractor for phase chunks
List<Chunk> chunks = JCasUtil.selectCovered(jCas, Chunk.class, sentence);
chunkingExtractors.add(new ChunkingExtractor("PhraseTag", this.phraseChunking, jCas, tokens, chunks));
// extract features for all tokens
int tokenIndex = -1;
int nChunkLabelsBefore = 2;
int nChunkLabelsAfter = 2;
int nPreviousClassifications = 2;
for (BaseToken token : tokens) {
++tokenIndex;
List<Feature> features = new ArrayList<Feature>();
// features from previous classifications
for (int i = nPreviousClassifications; i > 0; --i) {
int index = tokenIndex - i;
String previousOutcome = index < 0 ? "O" : outcomes.get(index);
features.add(new Feature("PreviousOutcome_" + i, previousOutcome));
}
// features from token attributes
features.addAll(this.tokenFeatureExtractor.extract(jCas, token));
// features from surrounding tokens
features.addAll(this.contextFeatureExtractor.extractWithin(jCas, token, sentence));
// features from ends of entities
features.addAll(endOfEntityFeatures.get(token));
// features from surrounding entity, phrase, etc. chunk-labels
for (ChunkingExtractor extractor : chunkingExtractors) {
features.addAll(extractor.extract(tokenIndex, nChunkLabelsBefore, nChunkLabelsAfter));
}
// features from semantic roles
features.addAll(predicateArgumentExtractor.extract(token));
// apply feature selection, if necessary
if (this.featureSelection != null) {
features = this.featureSelection.transform(features);
}
// if training, write to data file
if (this.isTraining()) {
String outcome = outcomes.get(tokenIndex);
// if it is an "O" down-sample it
if (outcome.equals("O")) {
if (rand.nextDouble() <= this.probabilityOfKeepingANegativeExample){
this.dataWriter.write(new Instance<String>(outcome, features));
}
}else{//for minority instances:
Instance<String> minorityInst = new Instance<String>(outcome, features);
this.dataWriter.write(minorityInst);
smote.addInstance(minorityInst);//add minority instances to SMOTE algorithm
}
}
// if predicting, add prediction to outcomes
else {
outcomes.add(this.classifier.classify(features));
}
}
// during prediction, convert chunk labels to events and add them to the CAS
if (!this.isTraining()) {
this.eventChunking.createChunks(jCas, tokens, outcomes);
}
}
if(this.isTraining() && this.smoteNumOfNeighbors >= 1){ //add synthetic instances to datawriter, if smote is selected
Iterable<Instance<String>> syntheticInsts = smote.populateMinorityClass();
for( Instance<String> sytheticInst: syntheticInsts){
this.dataWriter.write(sytheticInst);
}
}