def filterArticles(articles):
relevant_articles = {}
correct = [0] * (len(int2tags) - 1)
gold_num = [0] * (len(int2tags) - 1)
filtered_correct = [0] * (len(int2tags) - 1)
filtered_gold_num = [0] * (len(int2tags) - 1)
helper.load_constants()
print "Num incidents", len(incidents)
print "Num unfilitered articles", len(articles)
for incident_id in incidents.keys():
incident = incidents[incident_id]
if not 'citations' in incident:
continue
for citation_ind, citation in enumerate(incident['citations']):
saveFile = "../data/raw_data/" + incident_id + "_" + str(
citation_ind) + ".raw"
print "checking it for savefile", saveFile
if not saveFile in articles:
continue
article = tokenizer.tokenize(articles[saveFile])
ents = [incident[e.replace('-', '_')] for e in int2tags[1:]]
tags, cleanArticle = getTags(article, ents)
##Calculate scores for filitered and unflitered articles
for i in range(1, len(int2tags)):
correct[i - 1] += 1 if i in tags else 0
gold_num[i - 1] += ents[i -
1].strip().lower() not in ["unknown"]
if len(set(tags)) > 2: ##This is the filtering
for i in range(1, len(int2tags)):
filtered_correct[i - 1] += 1 if i in tags else 0
filtered_gold_num[i -
1] += ents[i -
1].strip().lower() not in [
"unknown"
]
#Store article in convenient format to writing to tagfile
relavant_article = {}
relavant_article['tokens'] = cleanArticle[:1000]
relavant_article['tags'] = tags
relavant_article['title'] = citation['Title']
relavant_article['ents'] = [cleanDelimiters(e) for e in ents]
relevant_articles[saveFile] = relavant_article
pickle.dump(relevant_articles, open('EMA_filtered_articles.2.p', 'wb'))
oracle_scores = [(correct[i] * 1. / gold_num[i],
int2tags[i + 1]) if gold_num[i] > 0 else 0
for i in range(len(correct))]
filtered_oracle_scores = [
(filtered_correct[i] * 1. / filtered_gold_num[i],
int2tags[i + 1]) if filtered_gold_num[i] > 0 else 0
for i in range(len(correct))
]
print "num articles is", len(relevant_articles)
print "oracle scores", oracle_scores
print "filtered_oracle_scores", filtered_oracle_scores
return relevant_articles
def filterArticles(articles):
relevant_articles = {}
correct = [0] * (len(int2tags) -1 )
gold_num = [0] * (len(int2tags)-1)
filtered_correct = [0] * (len(int2tags) -1 )
filtered_gold_num = [0] * (len(int2tags)-1)
helper.load_constants()
print "Num incidents", len(incidents)
print "Num unfilitered articles", len(articles)
for incident_id in incidents.keys():
incident = incidents[incident_id]
if not 'citations' in incident:
continue
for citation_ind, citation in enumerate(incident['citations']):
saveFile = "../data/raw_data/"+ incident_id+"_"+str(citation_ind)+".raw"
print "checking it for savefile", saveFile
if not saveFile in articles:
continue
article = tokenizer.tokenize(articles[saveFile])
ents = [incident[e.replace('-','_')] for e in int2tags[1:]]
tags, cleanArticle = getTags(article, ents)
##Calculate scores for filitered and unflitered articles
for i in range(1, len(int2tags) ):
correct[i-1] += 1 if i in tags else 0
gold_num[i-1] += ents[i-1].strip().lower() not in ["unknown"]
if len(set(tags)) > 2: ##This is the filtering
for i in range(1, len(int2tags) ):
filtered_correct[i-1] += 1 if i in tags else 0
filtered_gold_num[i-1] += ents[i-1].strip().lower() not in ["unknown"]
#Store article in convenient format to writing to tagfile
relavant_article = {}
relavant_article['tokens'] = cleanArticle[:1000]
relavant_article['tags'] = tags
relavant_article['title'] = citation['Title']
relavant_article['ents'] = [cleanDelimiters(e) for e in ents]
relevant_articles[saveFile] = relavant_article
pickle.dump(relevant_articles, open('EMA_filtered_articles.2.p', 'wb'))
oracle_scores = [(correct[i]*1./gold_num[i], int2tags[i+1]) if gold_num[i] > 0 else 0 for i in range(len(correct))]
filtered_oracle_scores = [(filtered_correct[i]*1./filtered_gold_num[i], int2tags[i+1]) if filtered_gold_num[i] > 0 else 0 for i in range(len(correct))]
print "num articles is", len(relevant_articles)
print "oracle scores", oracle_scores
print "filtered_oracle_scores", filtered_oracle_scores
return relevant_articles
def main(training_file, trained_model, previous_n, next_n, c, prune,
test_file):
helper.load_constants()
train_data, identifier = load_data(training_file)
test_data, test_ident = load_data(test_file)
## extract features
tic = time.clock()
print "get word_vocab"
num_words, word_vocab = get_word_vocab(train_data, prune)
print "feature extract for train"
trainX, trainY = get_feature_matrix_n(previous_n, next_n, train_data,
num_words, word_vocab,
helper.other_features)
print 'feature extract for test'
testX, testY = get_feature_matrix_n(previous_n, next_n, test_data,
num_words, word_vocab,
helper.other_features)
print time.clock() - tic
## train LR
print("training")
tic = time.clock()
clf = LogisticRegression(C=c, multi_class='multinomial', solver='lbfgs')
clf.fit(trainX, trainY)
print time.clock() - tic
print "predicting"
predictY = clf.predict(testX)
assert len(predictY) == len(testY)
print "evaluating"
evaluatePredictions(predictY, testY)
feature_list = (word_vocab.keys() + helper.other_features) * (
previous_n + next_n + 1) + word_vocab.keys() + ['previous_one'] * len(
tags) + ['previous_two'] * len(tags) + ['previous_three'
] * len(tags)
# getTopFeatures(clf,tags,feature_list)
if trained_model != "":
pickle.dump(
[clf, previous_n, next_n, word_vocab, helper.other_features],
open(trained_model, "wb"))
return [clf, previous_n, next_n, word_vocab, helper.other_features]
def main(training_file,trained_model,previous_n,next_n, c, prune, test_file):
helper.load_constants()
train_data, identifier = load_data(training_file)
test_data, test_ident = load_data(test_file)
## extract features
tic = time.clock()
print "get word_vocab"
num_words, word_vocab = get_word_vocab(train_data, prune)
print "feature extract for train"
trainX, trainY = get_feature_matrix_n(previous_n,next_n,train_data, num_words, word_vocab, helper.other_features)
print 'feature extract for test'
testX, testY = get_feature_matrix_n(previous_n, next_n, test_data, num_words, word_vocab, helper.other_features)
print time.clock()-tic
## train LR
print("training")
tic = time.clock()
clf = LogisticRegression(C=c, multi_class='multinomial', solver='lbfgs')
clf.fit(trainX,trainY)
print time.clock()-tic
print "predicting"
predictY = clf.predict(testX)
assert len(predictY) == len(testY)
print "evaluating"
evaluatePredictions(predictY, testY)
feature_list = (word_vocab.keys() + helper.other_features) * (previous_n+next_n+1) + word_vocab.keys() + ['previous_one'] * len(tags) + ['previous_two'] * len(tags)+ ['previous_three'] * len(tags)
# getTopFeatures(clf,tags,feature_list)
if trained_model != "":
pickle.dump([clf, previous_n,next_n, word_vocab,helper.other_features], open( trained_model, "wb" ) )
return [clf, previous_n,next_n, word_vocab,helper.other_features]
from train import load_data
import helper
import re, pdb, collections
import constants
import re
p = inflect.engine()
int2tags = [
'TAG'
] + constants.int2tags #since the constants file does not include the 'TAG' tag
NUM_ENTITIES = len(constants.int2tags)
tags2int = constants.tags2int
tags = range(len(int2tags))
helper.load_constants()
mode = constants.mode
CORRECT = collections.defaultdict(lambda: 0.)
GOLD = collections.defaultdict(lambda: 0.)
PRED = collections.defaultdict(lambda: 0.)
def splitBars(w):
return [q.strip() for q in w.split('|')]
# main loop
def main(trained_model,
testing_file,
viterbi,
import pickle
import inflect
import train_crf as crf
from train import load_data
import helper
import re, pdb, collections
import constants
import re
p = inflect.engine()
int2tags = ['TAG'] + constants.int2tags #since the constants file does not include the 'TAG' tag
tags2int = constants.tags2int
tags = range(len(int2tags))
helper.load_constants()
mode = constants.mode
# main loop
def main(trained_model,testing_file,viterbi,output_tags="output.tag", output_predictions="output.pred"):
test_data, identifier = load_data(testing_file)
evaluate = True
## extract features
if not "crf" in trained_model:
if not isinstance(trained_model, list):
clf, previous_n, next_n, word_vocab,other_features = pickle.load( open( trained_model, "rb" ) )
else:
clf, previous_n, next_n, word_vocab,other_features = trained_model
