def test_predicting_thrice(): trainCorpus1, trainCorpus2 = generateTestData(positiveCount=100,negativeCount=100,relTypes=1) trainCorpus3, testCorpus = generateTestData(positiveCount=100,negativeCount=100,relTypes=1) for doc in trainCorpus1.documents: for r in doc.relations: r.relationType = 'type1' for doc in trainCorpus2.documents: for r in doc.relations: r.relationType = 'type2' for doc in trainCorpus3.documents: for r in doc.relations: r.relationType = 'type3' testCorpus.removeRelations() classifier1 = kindred.RelationClassifier() classifier1.train(trainCorpus1) classifier2 = kindred.RelationClassifier() classifier2.train(trainCorpus2) classifier3 = kindred.RelationClassifier() classifier3.train(trainCorpus3) classifier1.predict(testCorpus) classifier2.predict(testCorpus) classifier3.predict(testCorpus) relations = [ r for doc in testCorpus.documents for r in doc.relations ] assert len(relations) == len(set(relations)), "Duplicate relations found in predictions"
def test_simpleRelationClassifier_triple():
trainCorpus, testCorpusGold = generateTestData(entityCount=3,
positiveCount=100,
negativeCount=100)
trainRelations = trainCorpus.getRelations()
assert len(trainRelations) == 50
for r in trainRelations:
assert len(r.entities) == 3
predictionCorpus = testCorpusGold.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier(entityCount=3)
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
predictedRelations = predictionCorpus.getRelations()
assert len(predictedRelations) == 50
for r in predictedRelations:
assert len(r.entities) == 3
f1score = kindred.evaluate(testCorpusGold,
predictionCorpus,
metric='f1score')
assert f1score == 1.0
def test_singleClassifier_triple():
trainCorpus, devCorpus = generateTestData(entityCount=3,
positiveCount=100,
negativeCount=100,
relTypes=1)
trainRelations = trainCorpus.getRelations()
assert len(trainRelations) == 50
for r in trainRelations:
assert len(r.entities) == 3
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier(entityCount=3)
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
predictedRelations = predictionCorpus.getRelations()
assert len(predictedRelations) == 50
for r in predictedRelations:
assert len(r.entities) == 3
f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score')
assert round(f1score, 3) == 1.0
def _featureBuilding(useBB3):
#trainData, testData = generateTestData(positiveCount=100,negativeCount=100)
if useBB3:
trainData = kindred.bionlpst.load('2016-BB3-event-train')
testData = kindred.bionlpst.load('2016-BB3-event-dev')
chosenFeatures = ["entityTypes", "unigramsBetweenEntities"]
else:
trainData = kindred.bionlpst.load('2016-SeeDev-binary-train')
testData = kindred.bionlpst.load('2016-SeeDev-binary-dev')
chosenFeatures = [
"entityTypes", "unigramsBetweenEntities", "bigrams",
"dependencyPathEdges", "dependencyPathEdgesNearEntities"
]
testData_TextAndEntities = [d.getTextAndEntities() for d in testData]
testData_Relations = [d.getRelations() for d in testData]
finalChosenFeatures = []
for threshold in range(1, 100, 2):
fthreshold = threshold / float(100)
classifier = kindred.RelationClassifier(useBuilder=False,
features=chosenFeatures,
threshold=fthreshold)
classifier.train(trainData)
predictedRelations = classifier.predict(testData_TextAndEntities)
evaluator = Evaluator()
precision, recall, f1score = evaluator.evaluate(testData_Relations,
predictedRelations,
metric='all',
display=False)
print(fthreshold, precision, recall, f1score)
def _nary():
trainData = kindred.BioNLPSTData(
'2016-SeeDev-binary-training_and_development')
testData = kindred.BioNLPSTData('2016-SeeDev-binary-test')
classifier_merge = kindred.RelationClassifier(naryMethod='merge')
classifier_merge.train(trainData)
predictedRelations_merge = classifier_merge.predict(testData)
kindred.saveST('BB3-predictions-merge/', testData,
predictedRelations_merge)
classifier_combined = kindred.RelationClassifier(naryMethod='combine')
classifier_combined.train(trainData)
predictedRelations_combined = classifier_combined.predict(testData)
kindred.saveST('BB3-predictions-combined/', testData,
predictedRelations_combined)
def test_doublelabels():
trainCorpus, devCorpus = generateTestData(positiveCount=100,
negativeCount=100,
relTypes=1)
for doc in trainCorpus.documents:
newRelations = [
kindred.Relation("anotherLabel", r.entities, r.argNames)
for r in doc.relations
]
doc.relations += newRelations
for doc in devCorpus.documents:
newRelations = [
kindred.Relation("anotherLabel", r.entities, r.argNames)
for r in doc.relations
]
doc.relations += newRelations
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score')
assert round(f1score, 3) == 1.0
def test_simpleRelationClassifier_emptyTrainCorpus():
trainCorpus, testCorpus = generateTestData(positiveCount=100,negativeCount=100)
for doc in trainCorpus.documents:
doc.entities = []
doc.relations = []
classifier = kindred.RelationClassifier()
with pytest.raises(RuntimeError) as excinfo:
classifier.train(trainCorpus)
assert excinfo.value.args == ('No candidate relations found in corpus for training. Does the corpus contain text and entity annotations with at least one sentence containing 2 entities.',)
def test_logisticregression_threshold(): trainCorpus, devCorpus = generateTestData(positiveCount=100,negativeCount=100,relTypes=1) predictionCorpus = devCorpus.clone() predictionCorpus.removeRelations() classifier = kindred.RelationClassifier(classifierType='LogisticRegression',threshold=0.3) classifier.train(trainCorpus) classifier.predict(predictionCorpus) f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score') assert round(f1score,3) == 1.0
def test_filterByEntityTypes_validTypes(): trainCorpus, devCorpus = generateTestData(positiveCount=100,negativeCount=100,relTypes=1) predictionCorpus = devCorpus.clone() predictionCorpus.removeRelations() classifier = kindred.RelationClassifier(features=["unigramsBetweenEntities","bigrams","dependencyPathEdges","dependencyPathEdgesNearEntities"]) classifier.train(trainCorpus) classifier.predict(predictionCorpus) f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score') assert round(f1score,3) == 1.0
def test_simpleRelationClassifier_emptyTestCorpus(): trainCorpus, testCorpus = generateTestData(positiveCount=100,negativeCount=100) for doc in testCorpus.documents: doc.entities = [] doc.relations = [] classifier = kindred.RelationClassifier() classifier.train(trainCorpus) classifier.predict(testCorpus) assert len(testCorpus.getRelations()) == 0
def test_simpleRelationClassifier_binary(): trainCorpus, testCorpusGold = generateTestData(positiveCount=100,negativeCount=100) predictionCorpus = testCorpusGold.clone() predictionCorpus.removeRelations() classifier = kindred.RelationClassifier() classifier.train(trainCorpus) classifier.predict(predictionCorpus) f1score = kindred.evaluate(testCorpusGold, predictionCorpus, metric='f1score') assert f1score == 1.0
def test_singleClassifier(): trainCorpus, devCorpus = generateTestData(positiveCount=100,negativeCount=100,relTypes=1) predictionCorpus = devCorpus.clone() predictionCorpus.removeRelations() classifier = kindred.RelationClassifier() classifier.train(trainCorpus) classifier.predict(predictionCorpus) f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score') assert round(f1score,3) == 1.0
def test_bionlp_BB3_classifier():
trainCorpus = kindred.bionlpst.load('2016-BB3-event-train')
devCorpus = kindred.bionlpst.load('2016-BB3-event-dev')
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score')
assert f1score > 0.4
def _bionlpst_seedev():
trainCorpus = kindred.bionlpst.load('2016-SeeDev-binary-train')
devCorpus = kindred.bionlpst.load('2016-SeeDev-binary-dev')
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score')
print("f1score:", f1score)
def test_bionlp_SeeDev_classifier():
trainCorpus = kindred.bionlpst.load('2016-SeeDev-binary-train')
devCorpus = kindred.bionlpst.load('2016-SeeDev-binary-dev')
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score')
assert round(f1score, 3) == 0.349
def test_predicting_duplicates(): trainCorpus, testCorpus = generateTestData(positiveCount=100,negativeCount=100,relTypes=1) testCorpus.removeRelations() classifier = kindred.RelationClassifier() classifier.train(trainCorpus) classifier.predict(testCorpus) classifier.predict(testCorpus) classifier.predict(testCorpus) relations = [ r for doc in testCorpus.documents for r in doc.relations ] assert len(relations) == len(set(relations)), "Duplicate relations found in predictions"
def _bionlpst_seedev(swap):
trainCorpus = kindred.bionlpst.load('2016-SeeDev-binary-train')
devCorpus = kindred.bionlpst.load('2016-SeeDev-binary-dev')
if swap:
trainCorpus,devCorpus = devCorpus,trainCorpus
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
scores = kindred.evaluate(devCorpus, predictionCorpus, metric='all')
print("seedev scores:",scores,swap)
def _SeeDevmini():
trainCorpus = kindred.bionlpst.load('2016-SeeDev-binary-train')
devCorpus = kindred.bionlpst.load('2016-SeeDev-binary-dev')
trainCorpus.documents = trainCorpus.documents[1:2]
devCorpus.documents = devCorpus.documents[:1]
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score')
assert round(f1score, 3) == 0.235
def _bionlpst_bb3(swap):
trainCorpus = kindred.bionlpst.load('2016-BB3-event-train')
devCorpus = kindred.bionlpst.load('2016-BB3-event-dev')
if swap:
trainCorpus,devCorpus = devCorpus,trainCorpus
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier(useBuilder=True)
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
scores = kindred.evaluate(devCorpus, predictionCorpus, metric='all')
print("bb3 scores:",scores,swap)
def _featureBuilding(useBB3):
#trainData, testData = generateTestData(positiveCount=100,negativeCount=100)
if useBB3:
trainData = kindred.bionlpst.load('2016-BB3-event-train')
testData = kindred.bionlpst.load('2016-BB3-event-dev')
useBuilder = True
else:
trainData = kindred.bionlpst.load('2016-SeeDev-binary-train')
testData = kindred.bionlpst.load('2016-SeeDev-binary-dev')
useBuilder = False
testData_TextAndEntities = [d.getTextAndEntities() for d in testData]
testData_Relations = [d.getRelations() for d in testData]
featureChoice = [
"entityTypes", "unigramsBetweenEntities", "bigrams",
"dependencyPathEdges", "dependencyPathEdgesNearEntities"
]
finalChosenFeatures = []
for stage in range(3):
bestF1, bestFeature = -1, -1
for feature in featureChoice:
chosenFeatures = finalChosenFeatures + [feature]
classifier = kindred.RelationClassifier(useBuilder=useBuilder,
features=chosenFeatures)
classifier.train(trainData)
predictedRelations = classifier.predict(testData_TextAndEntities)
evaluator = Evaluator()
f1score = evaluator.evaluate(testData_Relations,
predictedRelations,
metric='f1score',
display=False)
print(stage, feature, f1score)
if f1score > bestF1:
bestF1 = f1score
bestFeature = feature
featureChoice.remove(bestFeature)
finalChosenFeatures.append(bestFeature)
def _bionlpst_seedev_testSet():
trainData = kindred.bionlpst.load('2016-SeeDev-binary-train')
devData = kindred.bionlpst.load('2016-SeeDev-binary-dev')
testData = kindred.bionlpst.load('2016-SeeDev-binary-test')
trainAndDevData = trainData + devData
print("Starting training...")
classifier = kindred.RelationClassifier()
classifier.train(trainAndDevData)
print("Predicting training...")
predictedRelations = classifier.predict(
testData) #devData_TextAndEntities)
print("Saving...")
outDir = 'out.SeeDev'
kindred.save(testData,
'standoff',
outDir,
predictedRelations=predictedRelations)
def test_pickle():
trainCorpus, testCorpusGold = generateTestData(positiveCount=100,
negativeCount=100)
predictionCorpus = testCorpusGold.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
with tempfile.NamedTemporaryFile() as tempF:
with open(tempF.name, 'wb') as f:
pickle.dump(classifier, f)
with open(tempF.name, 'rb') as f:
classifier = pickle.load(f)
classifier.predict(predictionCorpus)
f1score = kindred.evaluate(testCorpusGold,
predictionCorpus,
metric='f1score')
assert f1score == 1.0
def _bionlpst_bb3_testSet():
trainData = kindred.bionlpst.load('2016-BB3-event-train')
devData = kindred.bionlpst.load('2016-BB3-event-dev')
testData = kindred.bionlpst.load('2016-BB3-event-test')
trainAndDevData = trainData + devData
print("Starting training...")
classifier = kindred.RelationClassifier(useBuilder=True)
#classifier = RelationClassifier(useBuilder=False)
classifier.train(trainAndDevData)
print("Predicting training...")
predictedRelations = classifier.predict(
testData) #devData_TextAndEntities)
print("Saving...")
outDir = 'out.BB3'
kindred.save(testData,
'standoff',
outDir,
predictedRelations=predictedRelations)
['Driver', 'Oncogene', 'Tumor_Suppressor'], [
args.outModel_Driver, args.outModel_Oncogene,
args.outModel_TumorSuppressor
]):
print("Building %s model" % relationType)
print(" Loading training")
goldDir = 'gold'
trainCorpus = kindred.load('standoff', args.inTrain)
for doc in trainCorpus.documents:
doc.relations = [
r for r in doc.relations if r.relationType == relationType
]
print(" Doing training")
features = "entityTypes,unigramsBetweenEntities,bigrams,dependencyPathEdges,dependencyPathEdgesNearEntities".split(
',')
threshold = thresholds[relationType]
classifier = kindred.RelationClassifier(
classifierType='LogisticRegression',
threshold=threshold,
features=features,
acceptedEntityTypes=[('cancer', 'gene')])
classifier.train(trainCorpus)
print(" Saving classifer")
with open(outModel, 'wb') as f:
pickle.dump(classifier, f)
print(" Output done!")
# avg_nn = 0
count = 0
iter_num = 200
print("-------------5 CLASSES------------")
while count < iter_num:
print("------", count, "------")
trainCorpus, devCorpus = Corpus.split(trainFraction=0.9)
predictionCorpus = devCorpus.clone()
predictionCorpus.removeRelations()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
svmf1score = kindred.evaluate(devCorpus,
predictionCorpus,
metric='f1score',
display=False)
print("svm:\t", svmf1score)
# avg_svm = avg_svm+svmf1score
classifier = kindred.RelationClassifier(classifierType='DCT')
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
dctf1score = kindred.evaluate(devCorpus,
predictionCorpus,
metric='f1score',
# Trim back to relation type of choice
for doc in train.documents:
doc.relations = [
r for r in doc.relations if r.relationType == args.reltype
]
for doc in gold.documents:
doc.relations = [
r for r in doc.relations if r.relationType == args.reltype
]
entityType = entityTypes[args.reltype]
entityCount = len(entityType)
classifier = kindred.RelationClassifier(
classifierType='LogisticRegression',
threshold=threshold,
entityCount=entityCount,
acceptedEntityTypes=[entityType])
classifier.train(train)
predictions = gold.clone()
predictions.removeRelations()
classifier.predict(predictions)
TP, FN, FP = 0, 0, 0
for goldDoc, testDoc in zip(gold.documents, predictions.documents):
goldTuples = [(r.relationType, tuple(r.entityIDs))
for r in goldDoc.relations]
testTuples = [(r.relationType, tuple(r.entityIDs))
help='Directory containing stand-off testing test')
args = parser.parse_args()
print("threshold\tprecision\trecall")
for threshold in np.arange(0, 1.01, 0.01):
trainCorpus = kindred.load('standoff', args.train)
testCorpus = kindred.load('standoff', args.test)
predCorpus = testCorpus.clone()
predCorpus.removeRelations()
parser = kindred.Parser(model='en_core_sci_sm')
parser.parse(trainCorpus)
parser.parse(testCorpus)
parser.parse(predCorpus)
classifier = kindred.RelationClassifier(
classifierType='LogisticRegression',
threshold=threshold,
acceptedEntityTypes=[('Chemical', 'Mutation')])
classifier.train(trainCorpus)
classifier.predict(predCorpus)
precision, recall, f1score = kindred.evaluate(testCorpus,
predCorpus,
metric='all',
display=False)
print("%f\t%f\t%f" % (threshold, precision, recall))
def get_rels_clean():
c.execute("SELECT gene, disease, relation FROM relationships")
return c.fetchall()
# Kindred
# 5 classes
trainCorpus = kindred.load(dataFormat='json', path='relation/db/1')
devCorpus = kindred.load(dataFormat='json', path='ner-dump')
predictionCorpus = devCorpus.clone()
classifier = kindred.RelationClassifier()
classifier.train(trainCorpus)
classifier.predict(predictionCorpus)
f1score = kindred.evaluate(devCorpus, predictionCorpus, metric='f1score')
print("5 CLASSES ---------------------")
for i in predictionCorpus.documents:
for j in (i.relations):
rel = Relationships(j.entities[0].text, j.entities[1].text,
j.relationType, i.text)
insert_rels(rel)
result = get_rels()
for i in result:
print(i)
for relationType, threshold, entityTypes in relationInfo:
print("Building %s model" % relationType)
print(" Loading training")
trainCorpus = kindred.load('standoff', args.inTrain)
for doc in trainCorpus.documents:
doc.relations = [
r for r in doc.relations if r.relationType == relationType
]
doc.relations = [
r for r in doc.relations if len(r.entities) == len(entityTypes)
]
print(" Doing training")
threshold = 0.5
classifier = kindred.RelationClassifier(
classifierType='LogisticRegression',
threshold=threshold,
entityCount=len(entityTypes),
acceptedEntityTypes=[entityTypes],
model='en_core_sci_sm')
classifier.train(trainCorpus)
print(" Saving classifer")
outModel = os.path.join(args.outDir, "%s.model" % relationType)
with open(outModel, 'wb') as f:
pickle.dump(classifier, f)
print(" Output done!")
