def decode(sents, goldtagseqs, postagseqs, info, weights) : #estfile, weightsfile):
labelset = ['B', 'I', 'O', '*']
tp_bi = 0
tp_o = 0
acc = 0.0
tot = 0
tot_rec_bi = 0
tot_rec_o = 0
tot_prec_bi = 0
tot_prec_o = 0
sys.stderr.write("total test sentences = " + str(len(sents)) + "\n")
for i in range(len(sents)):
sys.stderr.write(str(i) + "\r")
sent = sents[i]
postags = postagseqs[i]
tags = viterbi.execute(sent, labelset, postags, weights, info)
for j in range(len(tags)):
if tags[j] == goldtagseqs[i][j]:
acc += 1
if goldtagseqs[i][j] in ('B','I') and tags[j] in ('B','I'):
tp_bi += 1
elif goldtagseqs[i][j] == "O" and tags[j] == "O":
tp_o += 1
if goldtagseqs[i][j] in ('B', 'I'):
tot_rec_bi += 1
else:
tot_rec_o += 1
if tags[j] in ('B', 'I'):
tot_prec_bi += 1
else:
tot_prec_o += 1
print sent[j]+"\t"+postags[j]+"\t"+goldtagseqs[i][j]+"\t"+tags[j]
print
tot += len(tags)
sys.stderr.write("accuracy = " + str(acc/tot) + "\n")
sys.stderr.write("BI recall = " + str(tp_bi/tot_rec_bi) + "\n")
if tot_prec_bi > 0:
sys.stderr.write("BI precision = " + str(tp_bi/tot_prec_bi) + "\n")
sys.stderr.write("O recall = " + str(tp_o/tot_rec_o) + "\n")
if tot_prec_o > 0:
sys.stderr.write("O precision = " + str(tp_o/tot_prec_o) + "\n\n")
def run(sentset, labelset, postagseqs, vecs1, vecs2, num_iter, all_feats):
weights = init(all_feats)
weights_avg = init(all_feats)
for i in range(num_iter):
sys.stderr.write(str(i)+"\r")
for j in range(len(sentset)):
sent = sentset[j]
labelseq = labelset[j]
postagseq = postagseqs[j]
vec1 = vecs1[j]
vec2 = vecs2[j]
predseq, f = execute(sent, all_labels, postagseq, vec1, vec2, weights)
if labelseq != predseq:
update(weights, predseq, labelseq, postagseq, vec1, vec2, sent)
add_weights(weights_avg, weights)
for f in weights_avg.iterkeys():
weights_avg[f] /= num_iter*len(sentset)
print f, weights_avg[f]
return weights_avg
def run(sentset, labelset, postagseqs, vecs1, vecs2, num_iter, all_feats):
weights = init(all_feats)
weights_avg = init(all_feats)
for i in range(num_iter):
sys.stderr.write(str(i) + "\r")
for j in range(len(sentset)):
sent = sentset[j]
labelseq = labelset[j]
postagseq = postagseqs[j]
vec1 = vecs1[j]
vec2 = vecs2[j]
predseq, f = execute(sent, all_labels, postagseq, vec1, vec2,
weights)
if labelseq != predseq:
update(weights, predseq, labelseq, postagseq, vec1, vec2, sent)
add_weights(weights_avg, weights)
for f in weights_avg.iterkeys():
weights_avg[f] /= num_iter * len(sentset)
print f, weights_avg[f]
return weights_avg
def main(testfile, featsfile):
labelset = ['0', '1', '*']
test = codecs.open(testfile, 'r', 'utf-8')
feats = set([])
sents = []
tagseqs = []
postagseqs = []
vecs1 = []
vecs2 = []
contents = []
sent = []
tags = []
postags = []
vec1 = []
vec2 = []
content = []
while 1:
line = test.readline()
if not line:
break
line = line.strip()
if line == "":
sents.append(sent)
tagseqs.append(tags)
postagseqs.append(postags)
vecs1.append(vec1)
vecs2.append(vec2)
contents.append(content)
sent = []
tags = []
postags = []
vec1 = []
vec2 = []
content = []
continue
cline = line.split("\t")
word = cline[1].strip()
#tag = cline[13].strip()
tag = '1'
pos = cline[3].strip()
v1 = cline[10].strip()
v2 = cline[11].strip()
sent.append(word.strip())
tags.append(tag.strip())
postags.append(pos.strip())
vec1.append(v1.strip())
vec2.append(v2.strip())
content.append(cline)
test.close()
weights = {}
feats = open(featsfile, 'r')
while 1:
line = feats.readline()
if not line:
break
line = line.strip()
f, wt = line.split(' ')
weights[f] = float(wt)
feats.close()
acc = 0.0
tot = 0
for i in range(len(sents)):
sent = sents[i]
postags = postagseqs[i]
vec1 = vecs1[i]
vec2 = vecs2[i]
tags, f = viterbi.execute(sent, labelset, postags, vec1, vec2, weights)
for j in range(len(tags)):
print_line_withmodification(contents[i][j], tags[j])
if tags[j] == tagseqs[i][j]:
acc += 1
print
tot += len(tags)
def main(testfile, featsfile):
labelset = ['0', '1', '*']
test = codecs.open(testfile, 'r', 'utf-8')
feats = set([])
sents = []
tagseqs = []
postagseqs = []
vecs1 = []
vecs2 = []
contents = []
sent = []
tags = []
postags = []
vec1 = []
vec2 = []
content = []
while 1:
line = test.readline()
if not line:
break
line = line.strip()
if line == "":
sents.append(sent)
tagseqs.append(tags)
postagseqs.append(postags)
vecs1.append(vec1)
vecs2.append(vec2)
contents.append(content)
sent = []
tags = []
postags = []
vec1 = []
vec2 = []
content = []
continue
cline = line.split("\t")
word = cline[1].strip()
#tag = cline[13].strip()
tag = '1'
pos = cline[3].strip()
v1 = cline[10].strip()
v2 = cline[11].strip()
sent.append(word.strip())
tags.append(tag.strip())
postags.append(pos.strip())
vec1.append(v1.strip())
vec2.append(v2.strip())
content.append(cline)
test.close()
weights = {}
feats = open(featsfile, 'r')
while 1:
line = feats.readline()
if not line:
break
line = line.strip()
f, wt = line.split(' ')
weights[f] = float(wt)
feats.close()
acc = 0.0
tot = 0
for i in range(len(sents)):
sent = sents[i]
postags = postagseqs[i]
vec1 = vecs1[i]
vec2 = vecs2[i]
tags, f = viterbi.execute(sent, labelset, postags, vec1, vec2, weights)
for j in range(len(tags)):
print_line_withmodification(contents[i][j],tags[j])
if tags[j] == tagseqs[i][j]:
acc += 1
print
tot += len(tags)
