def _sentence_decoder(self, x):
rx = ut.repr_as_sents(x, self.w2v, self.max_len_doc_sents,
self.max_len_doc_sent_words, self.sent_separator,
self.word_separator, self.padding_val)
sx = x.split(self.sent_separator)
lsx = len(sx)
padding_req = max(0, self.max_len_doc_sents - lsx)
sent_weights = self.all_att_model.predict(np.array(
[rx]))[1][0][padding_req:]
relevant_weights = sorted(
sent_weights.argsort()[-self.topk_sentences:][::-1])
sent_summary = []
for k in relevant_weights:
sent_summary.append(sx[k].strip())
return "\n".join(sent_summary)
def _word_decoder(self, x):
rx = ut.repr_as_sents(x, self.w2v, self.max_len_doc_sents,
self.max_len_doc_sent_words, self.sent_separator,
self.word_separator, self.padding_val)
sents_x = x.split(self.sent_separator)
words_per_sent_x = [l.split() for l in sents_x]
words_x = self.word_separator.join(sents_x).split()
lsx = len(sents_x)
padding_sents = max(0, self.max_len_doc_sents - lsx)
padding_words = [
max(0, self.max_len_doc_sent_words - len(l))
for l in words_per_sent_x
]
sent_weights = self.all_att_model.predict(np.array(
[rx]))[1][0][padding_sents:]
word_weights = self.all_att_model.predict(np.array([rx]))[0][0]
word_weights = [
word_weights[i][padding_words[i]:]
for i in range(len(word_weights))
]
word_weights = [
word_weights[i] * sent_weights[i] for i in range(len(sent_weights))
]
word_summary = []
top_pos_words = ut.get_topk_maxs(word_weights, self.topk_words)
top_pos_words = sorted(top_pos_words)
h = {}
for i in range(len(top_pos_words)):
if top_pos_words[i][0] not in h:
h[top_pos_words[i][0]] = [top_pos_words[i][1]]
else:
h[top_pos_words[i][0]].append(top_pos_words[i][1])
for k in h:
for j in h[k]:
word_summary.append(words_per_sent_x[k][j])
word_summary.append(".")
return " ".join(word_summary)
