def load_data(coref):
data_path = os.path.join(path, 'MCTest')
if coref:
with open(os.path.join(data_path, 'mc500.train.coref'), 'rb') as f:
train_stories = clean(pickle.load(f))
with open(os.path.join(data_path, 'mc500.dev.coref'), 'rb') as f:
dev_stories = clean(pickle.load(f))
with open(os.path.join(data_path, 'mc500.test.coref'), 'rb') as f:
test_stories = clean(pickle.load(f))
else:
train_stories = clean(load_stories(
os.path.join(data_path, 'mc500.train.tsv'),
os.path.join(data_path, 'mc500.train.ans')))
dev_stories = clean(load_stories(
os.path.join(data_path, 'mc500.dev.tsv'),
os.path.join(data_path, 'mc500.dev.ans')))
test_stories = clean(load_stories(
os.path.join(data_path, 'mc500.test.tsv'),
os.path.join(data_path, 'mc500.test.ans')))
return train_stories, dev_stories, test_stories
def load_data(coref):
data_path = os.path.join(path, "MCTest")
if coref:
with open(os.path.join(data_path, "mc160.train.coref"), "rb") as f:
train_stories = clean(pickle.load(f))
with open(os.path.join(data_path, "mc160.dev.coref"), "rb") as f:
dev_stories = clean(pickle.load(f))
with open(os.path.join(data_path, "mc160.test.coref"), "rb") as f:
test_stories = clean(pickle.load(f))
else:
train_stories = clean(
load_stories(os.path.join(data_path, "mc160.train.tsv"), os.path.join(data_path, "mc160.train.ans"))
)
dev_stories = clean(
load_stories(os.path.join(data_path, "mc160.dev.tsv"), os.path.join(data_path, "mc160.dev.ans"))
)
test_stories = clean(
load_stories(os.path.join(data_path, "mc160.test.tsv"), os.path.join(data_path, "mc160.test.ans"))
)
return train_stories, dev_stories, test_stories
def process(path):
stories = load_stories(path + '.tsv', path + '.ans')
processed = []
filelist = []
try:
tmpdir = mkdtemp()
for i, story in enumerate(stories):
infile = os.path.join(tmpdir, str(i) + '.txt')
with open(infile, 'w') as f:
f.write(story.text)
filelist.append(infile)
filelist_filename = os.path.join(tmpdir, 'filelist')
with open(filelist_filename, 'w') as f:
for name in filelist:
f.write(name)
f.write(os.linesep)
subprocess.check_call([
'java', '-cp', '*', '-Xmx2g',
'edu.stanford.nlp.pipeline.StanfordCoreNLP',
'-annotators', 'tokenize,ssplit,pos,lemma,ner,parse,dcoref',
'-filelist', filelist_filename])
for i, story in enumerate(stories):
outfile = str(i) + '.txt.out'
with open(outfile) as f:
lines = f.readlines()
lines = [l.strip() for l in lines]
lines = [l for l in lines if len(l) > 0]
processed.append(Story(parse(story.text, lines), story.queries))
finally:
shutil.rmtree(tmpdir)
with open(path + '.coref', 'wb') as f:
pickle.dump(processed, f, 2)
"""
Runs a baseline test on a weakly supervised memory network.
"""
def clean(stories):
return [s for s in stories if len(s.queries) > 1]
def compute_accuracy(stories, model):
accuracy = 0
for story in stories:
score = model.predict_answer(story)
accuracy += score
return float(accuracy) / float(len(stories))
train_stories = clean(load_stories('MCTest/mc160.train.tsv','MCTest/mc160.train.ans'))
dev_stories = clean(load_stories('MCTest/mc160.dev.tsv','MCTest/mc160.dev.ans'))
test_stories = clean(load_stories('MCTest/mc160.test.tsv','MCTest/mc160.test.ans'))
all_stories = train_stories + test_stories + dev_stories
# initialize with all stories to get full vocab
model = WeakMemoryNetwork(300, 256, all_stories, timetags=True, word2vec=False)
print 'Training Accuracy prior to training: ', compute_accuracy(train_stories, model)
print 'Testing Accuracy prior to training: ', compute_accuracy(test_stories, model)
# Train for a certain number of epochs
count = 0
for i in range(30):
for story in train_stories + dev_stories:
