def train(notes, train_timex=True, train_event=True, train_rel=True):
# TODO: need to do some filtering of tokens
# TODO: experiment with the feature of the 4 left and right taggings. do we
# only utilize taggings for each pass or do we incorporate taggings in different passes?
timexLabels = [] # BIO labelings for tokens in text.
timexFeatures = []
eventLabels = [] # EVENT or O labelings for tokens in text.
eventFeatures = []
eventClassLabels = [] # event class labelings for tokens in text.
eventClassFeatures = []
tlinkLabels = [] # temporal relation labelings for enitity pairs.
tlinkFeatures = []
timexClassifier = None
timexVectorizer = None
eventClassifier = None
eventVectorizer = None
eventClassClassifier = None
eventClassVectorizer = None
tlinkClassifier = None
tlinkVectorizer = None
for i, note in enumerate(notes):
print "note: {}".format(i)
print "note path: ", note.note_path
if train_timex is True:
# extract features to perform BIO labeling for timexs
tmpLabels = note.get_timex_labels()
for label in tmpLabels: timexLabels += label
timexFeatures += features.extract_timex_feature_set(note, tmpLabels)
if train_event is True:
# extract features to perform EVENT or O labeling.
tmpLabels = note.get_event_labels()
for label in tmpLabels: eventLabels += label
eventFeatures += features.extract_event_feature_set(note, tmpLabels, timexLabels=note.get_timex_labels())
# extract features to perform event class labeling.
tmpLabels = note.get_event_class_labels()
for label in tmpLabels: eventClassLabels += label
eventClassFeatures += features.extract_event_class_feature_set(note, tmpLabels, note.get_event_labels(), timexLabels=note.get_timex_labels())
if train_rel is True:
# extract features to classify relations between temporal entities.
tlinkLabels += note.get_tlink_labels()
tlinkFeatures += features.extract_tlink_features(note)
# TODO: when predicting, if gold standard is provided evaluate F-measure for each of the steps
#print "VOCAB:"
#print features._voc
#print
if train_timex is True:
print "\tTRAINING TIMEX"
# train model to perform BIO labeling for timexs
timexClassifier, timexVectorizer = _trainTimex(timexFeatures, timexLabels, grid=True)
#timexClassifier, timexVectorizer = _trainTimex(timexFeatures, timexLabels, grid=False)
if train_event is True:
# train model to label as EVENT or O
# TODO: filter non-timex only?
eventClassifier, eventVectorizer = _trainEvent(eventFeatures, eventLabels, grid=True)
# eventClassifier, eventVectorizer = _trainEvent(eventFeatures, eventLabels, grid=False)
# train model to label as a class of EVENT
# TODO: filter event only?
# TODO: should we be training over all tokens or those that are just EVENTs?
eventClassClassifier, eventClassVectorizer = _trainEventClass(eventClassFeatures, eventClassLabels, grid=True)
#eventClassClassifier, eventClassVectorizer = _trainEventClass(eventClassFeatures, eventClassLabels, grid=False)
if train_rel is True:
# train model to classify relations between temporal entities.
# TODO: add features back in.
#tlinkClassifier, tlinkVectorizer = _trainTlink(tlinkFeatures, tlinkLabels, grid=False)
tlinkClassifier, tlinkVectorizer = _trainTlink(tlinkFeatures, tlinkLabels, grid=True)
# will be accessed later for dumping
models = {"TIMEX":timexClassifier,
"EVENT":eventClassifier,
"EVENT_CLASS":eventClassClassifier,
"TLINK":tlinkClassifier}
vectorizers = {"TIMEX":timexVectorizer,
"EVENT":eventVectorizer,
"EVENT_CLASS":eventClassVectorizer,
"TLINK":tlinkVectorizer}
return models, vectorizers
def predict(note, predict_timex=True, predict_event=True, predict_rel=True):
# TODO: try and correct the flattening on the lists. might just end up being redundent?
# TODO: refactor this code. a lot of it is redundant.
# TODO: need to do some filtering of tokens
# TODO: experiment with the feature of the 4 left and right taggings. do we
# only utilize taggings for each pass or do we incorporate taggings in different passes?
global _models
global _vects
global _models_loaded
if _models_loaded is False:
sys.exit("Models not loaded. Cannot predict")
# get tokenized text
tokenized_text = note.get_tokenized_text()
# will be new iob_labels
iob_labels = []
timexLabels = []
eventLabels = []
eventClassLabels = []
tlink_labels = []
# init the number of lines for timexlabels
# we currently do not know what they are.
# get the tokens into a flast list, these are ordered by
# appearance within the document
tokens = []
for line in tokenized_text:
timexLabels.append([])
eventLabels.append([])
eventClassLabels.append([])
iob_labels.append([])
tokens += tokenized_text[line]
timex_count = 2
if predict_timex is True:
timexClassifier = _models["TIMEX"]
timexVectorizer = _vects["TIMEX"]
# get the timex feature set for the tokens within the note.
timexFeatures = features.extract_timex_feature_set(note, timexLabels, predict=True)
# sanity check
assert len(tokens) == len(timexFeatures)
# predict over the tokens and the features extracted.
for t, f in zip(tokens, timexFeatures):
features.update_features(t, f, timexLabels)
X = timexVectorizer.transform([f])
Y = list(timexClassifier.predict(X))
timexLabels[t["sentence_num"] - 1].append({'entity_label':Y[0],
'entity_type':None if Y[0] == 'O' else 'TIMEX3',
'entity_id':"t"+str(timex_count)})
timex_count += 1
iob_labels[t["sentence_num"] - 1].append(timexLabels[t["sentence_num"] - 1][-1])
event_count = 2
event_class_count = 2
if predict_event is True:
eventClassifier = _models["EVENT"]
eventVectorizer = _vects["EVENT"]
eventClassClassifier = _models["EVENT_CLASS"]
eventClassVectorizer = _vects["EVENT_CLASS"]
# get the timex feature set for the tokens within the note.
# don't get iob labels yet, they are inaccurate. need to predict first.
eventFeatures = features.extract_event_feature_set(note, eventLabels, predict=True, timexLabels=timexLabels)
# sanity check
assert len(tokens) == len(eventFeatures)
# TODO: need to do some filter. if something is already labeled then just skip over it.
# predict over the tokens and the features extracted.
for t, f in zip(tokens, eventFeatures):
features.update_features(t, f, eventLabels)
X = eventVectorizer.transform([f])
Y = list(eventClassifier.predict(X))
eventLabels[t["sentence_num"] - 1].append({'entity_label':Y[0],
'entity_type':None if Y[0] == 'O' else 'EVENT',
'entity_id':"e" + str(event_count)})
event_count += 1
# get the timex feature set for the tokens within the note.
eventClassFeatures = features.extract_event_class_feature_set(note, eventClassLabels, eventLabels, predict=True, timexLabels=timexLabels)
# sanity check
assert len(tokens) == len(eventClassFeatures)
# predict over the tokens and the features extracted.
for t, f in zip(tokens, eventClassFeatures):
# updates labels
features.update_features(t, f, eventClassLabels)
X = eventClassVectorizer.transform([f])
Y = list(eventClassClassifier.predict(X))
eventClassLabels[t["sentence_num"] - 1].append({'entity_label':Y[0],
'entity_type':None if Y[0] == 'O' else 'EVENT',
'entity_id':'e' + str(event_class_count)})
event_class_count += 1
if iob_labels[t["sentence_num"] - 1][t["token_offset"]]["entity_type"] == None:
iob_labels[t["sentence_num"] - 1][t["token_offset"]] = eventClassLabels[t["sentence_num"] - 1][-1]
_totLabels = []
for l in eventClassLabels:
_totLabels += l
print "predicted ZERO events? : ", len(_totLabels) == len([l for l in _totLabels if l["entity_label"] != 'O'])
_totLabels = []
for l in timexLabels:
_totLabels += l
print "predicted ZERO timex? :", len(_totLabels) == len([l for l in _totLabels if l["entity_label"] != 'O'])
if predict_timex is True and predict_event is True and predict_rel is True:
tlinkVectorizer = _vects["TLINK"]
tlinkClassifier = _models["TLINK"]
note.set_tlinked_entities(timexLabels,eventClassLabels)
note.set_iob_labels(iob_labels)
print "PREDICT: getting tlink features"
f = features.extract_tlink_features(note)
X = tlinkVectorizer.transform(f)
tlink_labels = list(tlinkClassifier.predict(X))
entity_labels = [label for line in iob_labels for label in line]
original_offsets = note.get_token_char_offsets()
return entity_labels, original_offsets, tlink_labels, tokens
