-
Notifications
You must be signed in to change notification settings - Fork 45
/
relevance_ranking.py
222 lines (170 loc) · 6.67 KB
/
relevance_ranking.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
from random import randint
from scipy import sparse
from multiprocessing import Pool
from collections import Counter
import gc,string
import requests,time
import numpy
import re
from math import log
import logging
import timeout_decorator
from bs4 import BeautifulSoup,SoupStrainer
import pickle
from underthesea import pos_tag
from synonyms import generateVariants
from underthesea import pos_tag, ner, word_tokenize
from entity_linking import extractEntVariants
#w2v = KeyedVectors.load('resources/word2vec-200')
stopwords = open('resources/stopwords_small.txt').read().split('\n')
stopwords = set([w.replace(' ','_') for w in stopwords])
punct_set = set([c for c in string.punctuation]) | set(['“','”',"...","–","…","..","•",'“','”'])
def cos_sim(a, b):
return numpy.dot(a, b) / (numpy.linalg.norm(a) * numpy.linalg.norm(b))
def document_vector(doc):
vec = [w2v.wv[i] for i in doc]
return numpy.sum(vec,axis = 0)
def embedding_similarity(s1,s2):
s1 = list(set(s1.lower().split()))
s2 = list(set(s2.lower().split()))
s1 = [word for word in s1 if word in w2v.wv.vocab]
s2 = [word for word in s2 if word in w2v.wv.vocab]
if len(s1) == 0 or len(s2) == 0:
return 0
return cos_sim(document_vector(s1),document_vector(s2))
def generateNgram(paper, ngram = 2, deli = '_', rmSet = {}):
words = paper.split()
if len(words) == 1:
return ''
ngrams = []
for i in range(0,len(words) - ngram + 1):
block = words[i:i + ngram]
if not any(w in rmSet for w in block):
ngrams.append(deli.join(block))
return ngrams
def generatePassages(document,n):
passages = []
paragraphs = document.split('\n\n')
for para in paragraphs:
sentences = para.rsplit(' . ')
if len(sentences) <= 3:
passages.append(' '.join(sentences))
else:
for i in range(0,len(sentences) - n + 1):
passages.append(' '.join([sentences[i + j] for j in range(0,n) if '?' not in sentences[i + j]]))
return passages
def passage_score(q_ngrams,passage):
try:
passage = passage.lower()
p_unigram = set(generateNgram(passage,1,'_',punct_set | stopwords))
uni_score = len(p_unigram & q_ngrams['unigram'])
p_bigram = set(generateNgram(passage,2,'_',punct_set | stopwords))
p_trigram = set(generateNgram(passage,3,'_',punct_set | stopwords))
p_fourgram= set(generateNgram(passage,4,'_',punct_set))
bi_score = len(p_bigram & q_ngrams['bigram'])
tri_score = len(p_trigram & q_ngrams['trigram'])
four_score = len(p_fourgram & q_ngrams['fourgram'])
#emd_sim = embedding_similarity(' '.join(p_unigram),' '.join(q_ngrams['unigram']))
emd_sim = 0
return uni_score + bi_score*2 + tri_score*3 + four_score*4 + emd_sim*3
except:
return 0
def passage_score_wrap(args):
return passage_score(args[0],args[1])
def chunks(l, n):
for i in range(0, len(l), n):
yield l[i:i + n]
def get_entities(seq):
i = 0
chunks = []
seq = seq + ['O'] # add sentinel
types = [tag.split('-')[-1] for tag in seq]
while i < len(seq):
if seq[i].startswith('B'):
for j in range(i+1, len(seq)):
if seq[j].startswith('I') and types[j] == types[i]:
continue
break
chunks.append((types[i], i, j))
i = j
else:
i += 1
return chunks
def get_ner(text):
res = ner(text)
words = [r[0] for r in res]
tags = [t[3] for t in res]
chunks = get_entities(tags)
res = []
for chunk_type, chunk_start, chunk_end in chunks:
res.append(' '.join(words[chunk_start: chunk_end]))
return res
def keyword_extraction(question):
keywords = []
question = question.replace('_',' ')
if 'nhất' in question.lower():
keywords.append('nhất')
words = pos_tag(question)
for i in range(0,len(words)):
words[i] = (words[i][0].replace(' ','_'),words[i][1])
for token in words:
word = token[0]
pos = token[1]
if (pos in ['A','Ab']):
keywords += word.lower().split('_')
keywords = list(set(keywords))
keywords = [[w] for w in keywords]
ners = get_ner(question)
ners = [n.lower() for n in ners]
for ne in ners:
variants = extractEntVariants(ne)
keywords.append(variants)
return keywords
def isRelevant(text,keywords):
text = text.lower().replace('_',' ')
for words in keywords:
if not any(e for e in words if e in text):
return False
return True
def removeDuplicate(documents):
mapUnigram = {}
for doc in documents:
mapUnigram[doc] = generateNgram(doc.lower(),1,'_',punct_set | stopwords)
uniqueDocs = []
for i in range(0,len(documents)):
check = True
for j in range(0,len(uniqueDocs)):
check_doc = mapUnigram[documents[i]]
exists_doc = mapUnigram[uniqueDocs[j]]
overlap_score = len( set(check_doc) & set(exists_doc) )
if overlap_score >= 0.8 * len(set(check_doc)) or overlap_score >= 0.8 * len(set(exists_doc)):
check = False
if check:
uniqueDocs.append(documents[i])
return uniqueDocs
def rel_ranking(question,documents):
#Return ranked list of passages from list of documents
pool = Pool(4)
q_variants = generateVariants(question)
q_keywords = keyword_extraction(question)
q_ngrams = {'unigram': set(generateNgram(question.lower(),1,'_',punct_set | stopwords))
, 'bigram' : set([]), 'trigram': set([]), 'fourgram': set([])}
for q in q_variants:
q = q.lower()
q_ngrams['bigram'] = q_ngrams['bigram'] | set(generateNgram(q,2,'_',punct_set | stopwords))
q_ngrams['trigram'] = q_ngrams['trigram'] | set(generateNgram(q,3,'_',punct_set | stopwords))
q_ngrams['fourgram']= q_ngrams['fourgram'] | set(generateNgram(q,4,'_',punct_set))
documents = [d for d in documents if isRelevant(d,q_keywords)]
passages = [generatePassages(d,3) for d in documents]
passages = [j for i in passages for j in i]
passages = list(set(passages))
passages = [p for p in passages if isRelevant(p,q_keywords)]
p_scores = pool.map(passage_score_wrap,[(q_ngrams,p) for p in passages])
p_res = numpy.argsort([-s for s in p_scores])
relevantDocs = []
for i in range(0,len(passages)):
relevantDocs.append(passages[p_res[i]])
relevantDocs = removeDuplicate(relevantDocs)
pool.terminate()
del pool
return relevantDocs