Spaces:
Runtime error
Runtime error
mariamisoieva
commited on
Commit
•
ef89d5e
1
Parent(s):
dcc1589
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,496 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import tensorflow as tf
|
3 |
+
from transformers import TFGPT2LMHeadModel, GPT2Tokenizer
|
4 |
+
|
5 |
+
|
6 |
+
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
|
7 |
+
|
8 |
+
# add the EOS token as PAD token to avoid warnings
|
9 |
+
model = TFGPT2LMHeadModel.from_pretrained("gpt2", pad_token_id=tokenizer.eos_token_id)
|
10 |
+
import stanza
|
11 |
+
stanza.download('en')
|
12 |
+
import nltk
|
13 |
+
|
14 |
+
nltk.download('punkt')
|
15 |
+
nltk.download('wordnet')
|
16 |
+
savejson = True
|
17 |
+
indexing = True
|
18 |
+
import numpy as np
|
19 |
+
import pandas as pd
|
20 |
+
from nltk.stem import WordNetLemmatizer
|
21 |
+
from nltk.corpus import wordnet
|
22 |
+
from nltk.wsd import lesk
|
23 |
+
import stanza
|
24 |
+
import nltk
|
25 |
+
import collections
|
26 |
+
import itertools
|
27 |
+
import json
|
28 |
+
from collections import defaultdict
|
29 |
+
nlp = stanza.Pipeline()
|
30 |
+
lemmatizer = WordNetLemmatizer()
|
31 |
+
stories = pd.read_csv('ROCStories_winter2017 - ROCStories_winter2017 (1).csv')
|
32 |
+
import warnings
|
33 |
+
warnings.filterwarnings('ignore')
|
34 |
+
def defdict():
|
35 |
+
return defaultdict(list)
|
36 |
+
class Sentence:
|
37 |
+
def __init__(self, textid, sentencenum, sentence, vectors=None, tfidfs=None, sentenceVector=None, lemmatized=None, preds=None, args=None):
|
38 |
+
self.textid = textid
|
39 |
+
self.sentence = sentence
|
40 |
+
self.sentencenum = sentencenum
|
41 |
+
self.vectors = vectors
|
42 |
+
self.tfidfs = tfidfs
|
43 |
+
self.sentenceVector = sentenceVector
|
44 |
+
self.preds = preds
|
45 |
+
self.args = args
|
46 |
+
if lemmatized:
|
47 |
+
self.lemmatized = lemmatized
|
48 |
+
else:
|
49 |
+
self.lemmatize()
|
50 |
+
|
51 |
+
def lemmatize(self):
|
52 |
+
doc = nlp(self.sentence)
|
53 |
+
self.lemmatized = []
|
54 |
+
self.preds = []
|
55 |
+
ind = 0
|
56 |
+
self.args=[]
|
57 |
+
for i, dep_edge in enumerate(doc.sentences[0].dependencies):
|
58 |
+
if dep_edge[1] != 'punct':
|
59 |
+
self.lemmatized.append(dep_edge[2].lemma)
|
60 |
+
if dep_edge[1] == "root":
|
61 |
+
self.preds.append(dep_edge[2].lemma)
|
62 |
+
ind = i+1
|
63 |
+
for dep_edge in doc.sentences[0].dependencies:
|
64 |
+
if int(dep_edge[2].head) == ind and dep_edge[1] != 'punct': #dep_edge[1] not in x
|
65 |
+
self.args.append(lemmatizer.lemmatize(dep_edge[2].lemma.lower()))
|
66 |
+
self.doc = doc
|
67 |
+
|
68 |
+
def calculateVector(self):
|
69 |
+
if self.vectors and self.tfidfs:
|
70 |
+
self.sentenceVector = np.dot(self.tfidfs, self.vectors)
|
71 |
+
return self.sentenceVector
|
72 |
+
|
73 |
+
def getVector(self):
|
74 |
+
if self.sentenceVector is None:
|
75 |
+
self.calculateVector()
|
76 |
+
return self.sentenceVector
|
77 |
+
|
78 |
+
class Story:
|
79 |
+
def __init__(self, sentences, number):
|
80 |
+
self.sentences = sentences
|
81 |
+
self.number = number
|
82 |
+
def lemmatizedSents(self):
|
83 |
+
lemSents = []
|
84 |
+
for s in self.sentences:
|
85 |
+
lemSents.append(s.lemmatized)
|
86 |
+
return lemSents
|
87 |
+
storiesSentences = []
|
88 |
+
sentencesjsons =[]
|
89 |
+
def indexSents(sents):
|
90 |
+
ind = defaultdict(defdict)
|
91 |
+
for sc in sents:
|
92 |
+
for i, w in enumerate(sc.lemmatized):
|
93 |
+
ind[w][sc.textid].append((i, sc.sentencenum))
|
94 |
+
return ind
|
95 |
+
def indexCorpus():
|
96 |
+
sentences = []
|
97 |
+
# textid, sentencenum, sentence
|
98 |
+
for i, story in stories[:300].iterrows():
|
99 |
+
storiesSentences.append([])
|
100 |
+
# document = ""
|
101 |
+
print(i)
|
102 |
+
for sind, sent in enumerate(story[2:], start = 1):
|
103 |
+
sentence = Sentence(i, sind-1, sent)
|
104 |
+
# print(sent)
|
105 |
+
# print(i)
|
106 |
+
# print(sentence.sentencenum)
|
107 |
+
# document.join(sent)
|
108 |
+
storiesSentences[i].append(sentence)
|
109 |
+
sentences.append(sentence)
|
110 |
+
sentencesjsons.append(sentence.__dict__)
|
111 |
+
# storiesClasses.append(Story(storiesSentences[i],i))
|
112 |
+
# documents.append(document)
|
113 |
+
return indexSents(sentences)
|
114 |
+
if savejson:
|
115 |
+
index = indexCorpus()
|
116 |
+
#json.dump(sentencesjsons, open('filename.json', 'a'))
|
117 |
+
else:
|
118 |
+
sentencesjsons = json.load(open('filename.json'))
|
119 |
+
# if indexing:
|
120 |
+
# json.dump(index, open('index.json', 'w'))
|
121 |
+
# else:
|
122 |
+
# index = json.load(open('index.json'))
|
123 |
+
|
124 |
+
def searchByRequest(words):
|
125 |
+
sents = set()
|
126 |
+
dicts = []
|
127 |
+
keys = [] #story numbers
|
128 |
+
synonims = []
|
129 |
+
for i, w in enumerate(words):
|
130 |
+
synonims.append(set())
|
131 |
+
synonims[i].update([w])
|
132 |
+
for synset in wordnet.synsets(w):
|
133 |
+
synonims[i].update(synset.lemma_names())
|
134 |
+
# print(synonims)
|
135 |
+
stories = []
|
136 |
+
dictsForWords = []
|
137 |
+
storiesForWords = []
|
138 |
+
for i, w in enumerate(synonims):
|
139 |
+
dictsForWords.append([])
|
140 |
+
storiesForWords.append(set())
|
141 |
+
for synonim in w:
|
142 |
+
currrentDict = index[synonim]
|
143 |
+
if currrentDict:
|
144 |
+
dictsForWords[i].append(currrentDict)
|
145 |
+
storiesForWords[i].update(set(currrentDict.keys()))
|
146 |
+
paragraphs = set.intersection(*storiesForWords)
|
147 |
+
# print(paragraphs)
|
148 |
+
# print(dictsForWords)
|
149 |
+
# print(dicts)
|
150 |
+
sentencesClasses = set()
|
151 |
+
temporarySentencesByParagraphs = [[set()]*len(words)]*len(paragraphs)
|
152 |
+
for pi, p in enumerate(paragraphs):
|
153 |
+
temporarySentences = []
|
154 |
+
for wi, wordDictsList in enumerate(dictsForWords):
|
155 |
+
temporarySentences.append(set())
|
156 |
+
# print(wordDictsList)
|
157 |
+
for dictionary in wordDictsList:
|
158 |
+
if p in dictionary:
|
159 |
+
for sents in dictionary[p]:
|
160 |
+
# print(sents)
|
161 |
+
temporarySentences[wi].update([sents[1]])
|
162 |
+
# print(temporarySentences[wi])
|
163 |
+
# print(temporarySentences)
|
164 |
+
if wi>0 and len(words) > 1:
|
165 |
+
for i in range(wi):
|
166 |
+
for s in temporarySentences[wi]:
|
167 |
+
if s in temporarySentences[i]:
|
168 |
+
sentencesClasses.update([storiesSentences[p][s]])
|
169 |
+
# for sentence in sentencesClasses:
|
170 |
+
# print(sentence.lemmatized)
|
171 |
+
# print(sentence.sentence, sentence.textid, sentence.sentencenum)
|
172 |
+
return sentencesClasses
|
173 |
+
# m = searchByRequest(['play', 'fun', 'game'])
|
174 |
+
# m = searchByRequest(['present', 'Christmas', 'wake'])
|
175 |
+
def predIndex():
|
176 |
+
# stories = pd.read_csv('ROCStories_winter2017 - ROCStories_winter2017 (1).csv')
|
177 |
+
ind = defaultdict(defdict)
|
178 |
+
for i, story in enumerate(storiesSentences):
|
179 |
+
for j, sent in enumerate(story):
|
180 |
+
for s in sent.preds:
|
181 |
+
ind[s][i].append(j)
|
182 |
+
return ind
|
183 |
+
preds = predIndex()
|
184 |
+
def powC(subj):
|
185 |
+
c = 0
|
186 |
+
for k, v in preds[subj].items():
|
187 |
+
c += len(v)
|
188 |
+
return c
|
189 |
+
def powCons(s1, s2):
|
190 |
+
count = 0
|
191 |
+
for i in (preds[s1].keys() & preds[s2].keys()):
|
192 |
+
i1=0
|
193 |
+
i2=0
|
194 |
+
while i1 != len(preds[s1][i]) and i2 != len(preds[s2][i]): #for d in preds[s1][i]:
|
195 |
+
if preds[s1][i][i1] + 1 == preds[s2][i][i2]:
|
196 |
+
count += 1
|
197 |
+
i1 += 1
|
198 |
+
i2 += 1
|
199 |
+
elif preds[s1][i][i1] + 1 < preds[s2][i][i2]:
|
200 |
+
i1 += 1
|
201 |
+
else:
|
202 |
+
i2 += 1
|
203 |
+
return count
|
204 |
+
|
205 |
+
# print(powCons('decide', 'make'))
|
206 |
+
# print(powCons('know', 'buy'))
|
207 |
+
def synset_lesk(sent, word):
|
208 |
+
sent_tok = nltk.tokenize.word_tokenize(sent)
|
209 |
+
return lesk(sent_tok, word) #,pos
|
210 |
+
# comparison of wsd
|
211 |
+
# def wpsim():
|
212 |
+
# def wpsim_by_max():
|
213 |
+
def wpsim_lesk(word1, sent1, word2, sent2):
|
214 |
+
synset1 = lesk(sent1, word1)
|
215 |
+
# print(synset1.definition())
|
216 |
+
synset2 = lesk(sent2, word2)
|
217 |
+
# print(synset2.definition())
|
218 |
+
return synset1.wup_similarity(synset2)
|
219 |
+
x = ['punct', 'conj']
|
220 |
+
def args_of_pred(s):
|
221 |
+
return s.args
|
222 |
+
import math
|
223 |
+
alpha = 0.5
|
224 |
+
def FRelPred(sent1, sent2):
|
225 |
+
try:
|
226 |
+
p1 = sent1.preds[0]
|
227 |
+
p2 = sent2.preds[0]
|
228 |
+
if powCons(p1, p2) == 0:
|
229 |
+
return 0.0
|
230 |
+
return math.log2(powCons(p1, p2) / (powC(p1)*powC(p2)))
|
231 |
+
except:
|
232 |
+
return 0.2
|
233 |
+
def FRelArgs(s1, s2):
|
234 |
+
try:
|
235 |
+
args1 = args_of_pred(s1)
|
236 |
+
args2 = args_of_pred(s2)
|
237 |
+
# print(args1, args2)
|
238 |
+
sent_tok1 = s1.lemmatized#nltk.tokenize.word_tokenize(s1)
|
239 |
+
sent_tok2 = s2.lemmatized#nltk.tokenize.word_tokenize(s2)
|
240 |
+
# print(sent_tok1, sent_tok2)
|
241 |
+
sum1 = 0
|
242 |
+
sum2 = 0
|
243 |
+
max1 = 0
|
244 |
+
max2 = 0
|
245 |
+
wpsim = 0
|
246 |
+
for ni in args1:
|
247 |
+
synsetni = lesk(sent_tok1, ni) #pos
|
248 |
+
# print(synsetni, ni)
|
249 |
+
synsetnj = lesk(sent_tok2, args2[0])
|
250 |
+
# print(synsetnj, args2[0])
|
251 |
+
# if synsetni != None and synsetnj != None:
|
252 |
+
if not (synsetnj is None or synsetni is None):
|
253 |
+
max1 = synsetni.wup_similarity(synsetnj) #wp_sim(ni, args2[0])
|
254 |
+
# print(type(max1))
|
255 |
+
if max1 is None:
|
256 |
+
max1 = 0
|
257 |
+
for nj in args2[1:]:
|
258 |
+
synsetnj = lesk(sent_tok2, nj)
|
259 |
+
# print(synsetni, ni)
|
260 |
+
# print(synsetnj, nj)
|
261 |
+
# if synsetni != None and synsetnj != None:
|
262 |
+
if not (synsetnj is None or synsetni is None):
|
263 |
+
wpsim = synsetni.wup_similarity(synsetnj) #(ni, nj)
|
264 |
+
if wpsim is None:
|
265 |
+
wpsim = 0
|
266 |
+
if (not None in [wpsim, max1]) and wpsim > max1:
|
267 |
+
max1 = wpsim
|
268 |
+
# print(wpsim, max1)
|
269 |
+
sum1 += max1
|
270 |
+
# print(sum1)
|
271 |
+
|
272 |
+
for ni in args2:
|
273 |
+
synsetni = lesk(sent_tok2, ni)
|
274 |
+
synsetnj = lesk(sent_tok1, args1[0])
|
275 |
+
if not (synsetnj is None or synsetni is None):
|
276 |
+
max2 = synsetni.wup_similarity(synsetnj) #wp_sim(ni, args2[0])
|
277 |
+
if max2 is None:
|
278 |
+
max2 = 0
|
279 |
+
for nj in args1[1:]:
|
280 |
+
synsetnj = lesk(sent_tok1, nj)
|
281 |
+
if not (synsetnj is None or synsetni is None):
|
282 |
+
wpsim = synsetni.wup_similarity(synsetnj) #(ni, nj)
|
283 |
+
if wpsim is None:
|
284 |
+
wpsim = 0
|
285 |
+
if (not None in [wpsim, max2]) and wpsim > max2:
|
286 |
+
# if (wpsim is not None) and wpsim > max2:
|
287 |
+
max2 = wpsim
|
288 |
+
sum2 += max2
|
289 |
+
# print(len(args1))
|
290 |
+
# print(len(args2))
|
291 |
+
# print(sum1, sum2)
|
292 |
+
return 0.5*( (1/len(args1))*sum1 + (1/len(args2))*sum2 )
|
293 |
+
except:
|
294 |
+
return 0.2
|
295 |
+
def FRel(s1, s2):
|
296 |
+
return alpha*FRelPred(s1, s2) + (1-alpha)*FRelArgs(s1, s2)
|
297 |
+
def hac(foundSentences, length=2):
|
298 |
+
R=0.1
|
299 |
+
twoSentenceClusters = []
|
300 |
+
numfound = len(foundSentences)
|
301 |
+
|
302 |
+
sentencePairs = []
|
303 |
+
frelijs = []
|
304 |
+
ind = 0
|
305 |
+
maxind = 0
|
306 |
+
maxval = 0
|
307 |
+
for i in itertools.permutations(foundSentences, 2):
|
308 |
+
if i[0].textid != i[1].textid:
|
309 |
+
frelij = FRel(i[0], i[1])
|
310 |
+
if frelij > R:
|
311 |
+
sentencePairs.append(list(i))
|
312 |
+
frelijs.append(frelij)
|
313 |
+
if ind != 0:
|
314 |
+
if frelij > maxval:
|
315 |
+
maxval=maxval
|
316 |
+
maxind = ind
|
317 |
+
ind += 1
|
318 |
+
else:
|
319 |
+
ind=1
|
320 |
+
maxval = frelij
|
321 |
+
maxind = 0
|
322 |
+
|
323 |
+
# print(sentencePairs)
|
324 |
+
maxvalThree = 0
|
325 |
+
maxSentsThree = []
|
326 |
+
threeSentsCluster = set()
|
327 |
+
for pairind, pair in enumerate(sentencePairs):
|
328 |
+
for sent in foundSentences:
|
329 |
+
if sent.textid != pair[0].textid and sent.textid != pair[1].textid:
|
330 |
+
frelij = FRel(sent, i[0])
|
331 |
+
if frelij > R:
|
332 |
+
threeSentsCluster.add(tuple([sent]+pair))
|
333 |
+
current = (frelijs[maxind]+frelij)/2 > maxvalThree
|
334 |
+
if current > maxvalThree:
|
335 |
+
maxvalThree = current
|
336 |
+
maxSentsThree = [sent]+pair
|
337 |
+
frelij = FRel(i[1], sent)
|
338 |
+
if frelij > R:
|
339 |
+
threeSentsCluster.add(tuple(pair+[sent]))
|
340 |
+
current = (frelijs[maxind]+frelij)/2 > maxvalThree
|
341 |
+
if current > maxvalThree:
|
342 |
+
maxvalThree = current
|
343 |
+
maxSentsThree = pair+[sent]
|
344 |
+
# print(sentencePairs)
|
345 |
+
# print(threeSentsCluster)
|
346 |
+
# for pair in sentencePairs:
|
347 |
+
# print(pair[0].sentence, pair[1].sentence)
|
348 |
+
# for cluster in threeSentsCluster:
|
349 |
+
# print(cluster[0].sentence, cluster[1].sentence, cluster[2].sentence)
|
350 |
+
# print([sentencePairs[maxind],maxSentsThree])
|
351 |
+
if len(sentencePairs) >=1:
|
352 |
+
return [sentencePairs[maxind],maxSentsThree]#sentencePairs + list(threeSentsCluster)
|
353 |
+
else:
|
354 |
+
return []
|
355 |
+
# print(FRelPred('David noticed he had put on a lot of weight recently.',
|
356 |
+
# 'He examined his habits to try and figure out the reason.'))
|
357 |
+
# # 'After a few weeks, he started to feel much better.'))
|
358 |
+
# print(FRel('David noticed he had put on a lot of weight recently.',
|
359 |
+
# 'He examined his habits to try and figure out the reason.'))
|
360 |
+
# print(FRel('He decided to buy a pair of khakis.', 'The pair he bought fit him perfectly.'))
|
361 |
+
from sklearn.feature_extraction.text import TfidfVectorizer
|
362 |
+
from gensim.models.word2vec import Word2Vec
|
363 |
+
import gensim.downloader
|
364 |
+
def tfidfTokenizer(x):
|
365 |
+
return [w for words in [s.lemmatized for s in x] for w in words]
|
366 |
+
def preprocess(x):
|
367 |
+
return x
|
368 |
+
# tfidfvectorizer = TfidfVectorizer(tokenizer=tfidfTokenizer, preprocessor=preprocess, use_idf=True)
|
369 |
+
# tfidfvectorizer_vectors = tfidfvectorizer.fit_transform(storiesSentences) #(map(lambda x: [ s.lemmatized for s in np.array(x).flatten() ] ))
|
370 |
+
# wvModel = gensim.downloader.load("word2vec-google-news-300")
|
371 |
+
# feature_names = tfidfvectorizer.get_feature_names()
|
372 |
+
def setVectors(stories):
|
373 |
+
for doc in stories:
|
374 |
+
for sentence in doc:
|
375 |
+
vectors = []
|
376 |
+
for lemma in sentence.lemmatized:
|
377 |
+
try:
|
378 |
+
vectors.append(wvModel[lemma])
|
379 |
+
except:
|
380 |
+
vectors.append([0]*300)
|
381 |
+
sentence.vectors = vectors
|
382 |
+
# setVectors(storiesSentences)
|
383 |
+
def setTfIdfs(documents):
|
384 |
+
for i, doc in enumerate(documents):
|
385 |
+
feature = tfidfvectorizer_vectors[i,:].nonzero()[1]
|
386 |
+
tfidfs = zip(feature, [tfidfvectorizer_vectors[i, x] for x in feature])
|
387 |
+
tfidfsbyword = dict()
|
388 |
+
for w,s in [(feature_names[j], s) for (j, s) in tfidfs]:
|
389 |
+
tfidfsbyword[w] = s
|
390 |
+
for sent in doc:
|
391 |
+
tfidfs = []
|
392 |
+
for lemma in sent.lemmatized:
|
393 |
+
tfidfs.append(tfidfsbyword[lemma])
|
394 |
+
sent.tfidfs = tfidfs
|
395 |
+
sent.calculateVector()
|
396 |
+
# setTfIdfs(storiesSentences)
|
397 |
+
# m = searchByRequest(['wake', 'present', 'Christmas'])
|
398 |
+
# for sent in m:
|
399 |
+
# print(sent.lemmatized)
|
400 |
+
def generate(words):
|
401 |
+
m = searchByRequest(words)
|
402 |
+
return hac(m)
|
403 |
+
|
404 |
+
def generate_and_choose(input_text):
|
405 |
+
input_ids=tokenizer.encode(input_text,return_tensors='tf')
|
406 |
+
beam_outputs=model.generate(input_ids,max_length=100,num_return_sequences=3,num_beams=3,no_repeat_ngram_size=2,early_stopping=True)
|
407 |
+
return_list = []
|
408 |
+
for i, beam_output in enumerate(beam_outputs):
|
409 |
+
# print(beam_output)
|
410 |
+
return_list.append(tokenizer.decode(beam_output, skip_special_tokens=True,clean_up_tokenization_spaces=True))
|
411 |
+
outputs_coherences = []
|
412 |
+
for i, text in enumerate(return_list):
|
413 |
+
sentencesTokenized = nltk.sent_tokenize(text)
|
414 |
+
coherence_cur = 0
|
415 |
+
length = len(sentencesTokenized)
|
416 |
+
for s in range(length-1):
|
417 |
+
coherence_cur += FRel(sentencesTokenized[s], sentencesTokenized[s+1])
|
418 |
+
if length == 1:
|
419 |
+
length += 1
|
420 |
+
outputs_coherences.append(coherence_cur / (length-1))
|
421 |
+
index_of_max = outputs_coherences.index(max(outputs_coherences))
|
422 |
+
return return_list[index_of_max]
|
423 |
+
|
424 |
+
def greedy_generate(inp):
|
425 |
+
input_ids = tokenizer.encode(inp, return_tensors='tf')
|
426 |
+
greedy_output = model.generate(input_ids, pad_token_id=tokenizer.encode('.')[0], eos_token_id=tokenizer.encode('.')[0])
|
427 |
+
return tokenizer.decode(greedy_output[0], skip_special_tokens=True)
|
428 |
+
|
429 |
+
def with_sampling(input_ids):
|
430 |
+
tf.random.set_seed(0)
|
431 |
+
# activate sampling and deactivate top_k by setting top_k sampling to 0
|
432 |
+
sample_output = model.generate(
|
433 |
+
input_ids,
|
434 |
+
do_sample=True,
|
435 |
+
max_length=50,
|
436 |
+
top_k=0,
|
437 |
+
temperature=0.7)
|
438 |
+
return tokenizer.decode(sample_output[0], skip_special_tokens=True)
|
439 |
+
|
440 |
+
def with_top_k_sampling(input_ids):
|
441 |
+
tf.random.set_seed(0)
|
442 |
+
sample_output = model.generate(
|
443 |
+
input_ids,
|
444 |
+
do_sample=True,
|
445 |
+
max_length=50,
|
446 |
+
top_k=50)
|
447 |
+
return tokenizer.decode(sample_output[0], skip_special_tokens=True)
|
448 |
+
|
449 |
+
def with_nucleus_sampling(input_ids):
|
450 |
+
tf.random.set_seed(0)
|
451 |
+
# set top_k = 50 and set top_p = 0.95 and num_return_sequences = 3
|
452 |
+
sample_outputs = model.generate(
|
453 |
+
input_ids,
|
454 |
+
do_sample=True,
|
455 |
+
max_length=50,
|
456 |
+
top_k=50,
|
457 |
+
top_p=0.95,
|
458 |
+
num_return_sequences=3)
|
459 |
+
return_list = []
|
460 |
+
for i, beam_output in enumerate(sample_outputs):
|
461 |
+
# print(beam_output)
|
462 |
+
return_list.append(tokenizer.decode(beam_output, skip_special_tokens=True,clean_up_tokenization_spaces=True))
|
463 |
+
outputs_coherences = []
|
464 |
+
for i, text in enumerate(return_list):
|
465 |
+
sentencesTokenized = nltk.sent_tokenize(text)
|
466 |
+
coherence_cur = 0
|
467 |
+
length = len(sentencesTokenized)
|
468 |
+
for s in range(length-1):
|
469 |
+
coherence_cur += FRel(sentencesTokenized[s], sentencesTokenized[s+1])
|
470 |
+
if length == 1:
|
471 |
+
length += 1
|
472 |
+
outputs_coherences.append(coherence_cur / (length-1))
|
473 |
+
index_of_max = outputs_coherences.index(max(outputs_coherences))
|
474 |
+
return return_list[index_of_max]
|
475 |
+
|
476 |
+
def generation_method(decoding_algorithm,input_text):
|
477 |
+
input_ids=tokenizer.encode(input_text,return_tensors='tf')
|
478 |
+
if decoding_algorithm=="Beam search":
|
479 |
+
return generate_and_choose(input_text)
|
480 |
+
elif decoding_algorithm=="Greedy search":
|
481 |
+
return greedy_generate(input_text)
|
482 |
+
elif decoding_algorithm=="With sampling":
|
483 |
+
return with_sampling(input_ids)
|
484 |
+
elif decoding_algorithm=="With top k sampling":
|
485 |
+
return with_top_k_sampling(input_ids)
|
486 |
+
elif decoding_algorithm=="With nucleus sampling":
|
487 |
+
return with_nucleus_sampling(input_ids)
|
488 |
+
|
489 |
+
import gradio as gr
|
490 |
+
in1 = gr.inputs.Dropdown(choices=["Beam search", "Greedy search", "With sampling","With top k sampling", "With nucleus sampling"])
|
491 |
+
in2 = gr.inputs.Textbox()
|
492 |
+
iface = gr.Interface(fn=generation_method,
|
493 |
+
inputs=[in1,in2],
|
494 |
+
outputs="text").launch(debug=True)
|
495 |
+
|
496 |
+
iface.launch()
|