class WordEmbeddingAvgVectorConstructor:
def __init__(self):
self.data_accessor = CorpusAccessor()
if os.path.isfile(WORD_EMBEDDING_MODEL_PATH):
print('[INFO] loading word embedding model...')
self.word_embedding_model = word2vec.Word2Vec.load(
WORD_EMBEDDING_MODEL_PATH)
def convert_avg_vector(self, line):
"""
文を文中の各単語の平均ベクトルに変換
"""
if self.word_embedding_model is None:
raise ValueError("there is not word embedding model")
wakati_line = text_processor.wakati(line).split()
word_vectors = np.array([
self.word_embedding_model.__dict__['wv'][word]
for word in wakati_line
])
return np.average(word_vectors, axis=0)
def sentence_to_word_embedding_avg_vector(self, ncode):
"""
小説本文とあらすじ文の各文を、文中における各単語の分散表現の平均ベクトルに変換する
データは文番号をkey、文ベクトルをvalueとする辞書で保存される
[1: 文ベクトル, 2: 文ベクトル, ... , n: 文ベクトル]
"""
print('[PROCESS NCODE]: {}'.format(ncode))
contents_file_path = os.path.join(
WORD_EMBEDDING_AVG_VECTOR_CONTENTS_PATH, ncode + '.txt')
if os.path.isfile(contents_file_path):
return
contents_lines = self.data_accessor.get_contents_lines(ncode)
synopsis_lines = self.data_accessor.get_synopsis_lines(ncode)
if not contents_lines or not synopsis_lines:
return
# 本文各文のベクトル化
contents_line_vectors = dict()
for line_idx, line in enumerate(contents_lines):
if line_idx % 50 == 0:
print('contents progress: {:.1f}%'.format(
line_idx / len(contents_lines) * 100))
vector = self.convert_avg_vector(line)
contents_line_vectors[line_idx] = vector
# データの保存
print('[INFO] saving data: {}'.format(ncode))
with open(contents_file_path, 'wb') as cf:
joblib.dump(contents_line_vectors, cf, compress=3)
def construct(self):
for i, ncode in enumerate(self.data_accessor.ncodes):
print('[INFO] num of constructed data: {}'.format(i))
self.sentence_to_word_embedding_avg_vector(ncode)
class WordEmbeddingVectorConstructor:
def __init__(self):
self.data_accessor = CorpusAccessor()
if os.path.isfile(WORD_EMBEDDING_MODEL_PATH):
print('[INFO] loading word embedding model...')
self.word_embedding_model = word2vec.Word2Vec.load(
WORD_EMBEDDING_MODEL_PATH)
def convert_word_embedding_vectors(self, sentence):
"""
文を文中の単語の文さんベクトルのリストに変換
"""
wakati_line = text_processor.wakati(sentence).split()
return [
self.word_embedding_model.__dict__['wv'][word]
for word in wakati_line
]
def sentence_to_word_embedding_vectors(self, ncode):
"""
小説本文を、文中における各単語の分散表現のベクトルのリストに変換する
データは文番号をkey、文ベクトルをvalueとする辞書で保存される
[1: tensor, 2: tensor, ... , n: tensor]
"""
print('[PROCESS NCODE]: {}'.format(ncode))
contents_lines = self.data_accessor.get_contents_lines(ncode)
synopsis_lines = self.data_accessor.get_synopsis_lines(ncode)
if not contents_lines or not synopsis_lines:
return
# 本文各文のベクトル化
contents_line_vectors = dict()
for line_idx, line in enumerate(contents_lines):
if line_idx % 50 == 0:
print('contents progress: {:.1f}%'.format(
line_idx / len(contents_lines) * 100))
tensor = self.convert_word_embedding_vectors(line)
contents_line_vectors[line_idx] = tensor
# データの保存
contents_file_path = os.path.join(WORD_EMBEDDING_VECTORS_CONTENTS_PATH,
ncode + '.txt')
print('[INFO] saving data: {}'.format(ncode))
with open(contents_file_path, 'wb') as cf:
joblib.dump(contents_line_vectors, cf, compress=3)
def construct(self):
"""
全小説のデータを構築する
"""
for i, ncode in enumerate(self.data_accessor.ncodes):
print('[INFO] num of constructed data: {}'.format(i))
self.sentence_to_word_embedding_vectors(ncode)
class WordIndexesConstructor():
def __init__(self):
self.data_accessor = CorpusAccessor()
if os.path.isfile(WORD_EMBEDDING_MODEL_PATH):
self.word_embedding_model = word2vec.Word2Vec.load(WORD_EMBEDDING_MODEL_PATH)
def convert_index_list(self, line):
if self.word_embedding_model is None:
raise ValueError("there is not word embedding model")
words = text_processor.wakati(line).split()
index_list = [self.word_embedding_model.wv.vocab[word].index + 1 for word in words]
return index_list
def sentence_to_word_indexes(self, ncode):
"""
小説本文各文を、文中の各単語のインデックスのリストに変換する
データは文番号をkey、インデックスのリストをvalueとする辞書で保存される
[1: list, 2: list, ... , n: list]
"""
print('[PROCESS NCODE]: {}'.format(ncode))
contents_file_path = os.path.join(WORD_INDEXES_CONTENTS_PATH, ncode + '.txt')
if os.path.isfile(contents_file_path):
return
contents_lines = self.data_accessor.get_contents_lines(ncode)
synopsis_lines = self.data_accessor.get_synopsis_lines(ncode)
if not contents_lines or not synopsis_lines:
return
index_data = dict()
for line_idx, line in enumerate(contents_lines):
if line_idx % 50 == 0:
print('contents progress: {:.1f}%'.format(line_idx / len(contents_lines) * 100))
index_list = self.convert_index_list(line)
index_data[line_idx] = index_list
# データの保存
print('[INFO] saving data: {}'.format(ncode))
with open(contents_file_path, 'wb') as cf:
joblib.dump(index_data, cf, compress=3)
def construct(self):
"""
全小説のデータを構築する
"""
for i, ncode in enumerate(self.data_accessor.ncodes):
print('[INFO] num of constructed data: {}'.format(i))
self.sentence_to_word_indexes(ncode)
def generate(importance,
ncode,
position=False,
serif=False,
person=False,
sentence_length=False):
corpus_accessor = CorpusAccessor()
genre = corpus_accessor.get_genre(ncode)
if len(genre) == 0:
print('non genre')
return
s = LSTMSummarizer()
supplier = LSTMVectorSupplier(genre,
importance,
use_data_of_position_of_sentence=position,
use_data_of_is_serif=serif,
use_data_of_is_include_person=person,
use_data_of_sentence_length=sentence_length)
s.set_supplier(supplier)
s.set_trained_model()
print(s.generate(ncode=ncode))
def __init__(self):
self.data_accessor = CorpusAccessor()
if os.path.isfile(WORD_EMBEDDING_MODEL_PATH):
print('[INFO] loading word embedding model...')
self.word_embedding_model = word2vec.Word2Vec.load(
WORD_EMBEDDING_MODEL_PATH)
import numpy as np
from util.corpus_accessor import CorpusAccessor
corpus_accessor = CorpusAccessor()
def generate(ncode):
"""
正解データのスコアの大きい順に、参照あらすじと近い文字数のあらすじを生成する
"""
contents_lines = corpus_accessor.get_contents_lines(ncode)
synopsis_lines = corpus_accessor.get_synopsis_lines(ncode)
if not contents_lines or not synopsis_lines:
return
contents_lines = np.array(contents_lines)
# 参照あらすじの長さ
ref_length = len(''.join(synopsis_lines))
# 最低文数
min_sentence_count = 1
random_line_indexes = np.random.permutation(np.arange(len(contents_lines)))
hyp = contents_lines[random_line_indexes[:min_sentence_count]]
for sentence_index in random_line_indexes[min_sentence_count:]:
if len(''.join(np.append(hyp, contents_lines[sentence_index]))) <= ref_length:
hyp = np.append(hyp, contents_lines[sentence_index])
else:
break
return ''.join(hyp)
def multi_generate(importance, start, end):
"""
複数作品まとめて確認したいとき
"""
corpus_accessor = CorpusAccessor()
output_file_path = 'result_start_' + str(start) + '_end_' + str(
end) + '.txt'
file = open(output_file_path, 'w')
love_story_s = LSTMSummarizer()
love_story_supplier = LSTMVectorSupplier(
'love_story',
importance,
use_data_of_position_of_sentence=True,
use_data_of_is_serif=True,
use_data_of_is_include_person=True,
use_data_of_sentence_length=True)
love_story_s.set_supplier(love_story_supplier)
love_story_s.set_trained_model()
fantasy_s = LSTMSummarizer()
fantasy_supplier = LSTMVectorSupplier(
'fantasy',
importance,
use_data_of_position_of_sentence=True,
use_data_of_is_serif=True,
use_data_of_is_include_person=True,
use_data_of_sentence_length=True)
fantasy_s.set_supplier(fantasy_supplier)
fantasy_s.set_trained_model()
literature_s = LSTMSummarizer()
literature_supplier = LSTMVectorSupplier(
'literature',
importance,
use_data_of_position_of_sentence=True,
use_data_of_is_serif=True,
use_data_of_is_include_person=True,
use_data_of_sentence_length=True)
literature_s.set_supplier(literature_supplier)
literature_s.set_trained_model()
sf_s = LSTMSummarizer()
sf_supplier = LSTMVectorSupplier('sf',
importance,
use_data_of_position_of_sentence=True,
use_data_of_is_serif=True,
use_data_of_is_include_person=True,
use_data_of_sentence_length=True)
sf_s.set_supplier(sf_supplier)
sf_s.set_trained_model()
# sys.setrecursionlimit(20000)
rouge = Rouge()
for i, ncode in enumerate(corpus_accessor.exist_ncodes[start:end]):
print('processed ncode count: ', i)
genre = corpus_accessor.get_genre(ncode)
if len(genre) == 0:
print('non genre')
continue
ref = ''.join(corpus_accessor.get_synopsis_lines(ncode))
synopsis = ''
if genre == 'love_story':
synopsis = love_story_s.generate(ncode)
elif genre == 'fantasy':
synopsis = fantasy_s.generate(ncode)
elif genre == 'literature':
synopsis = literature_s.generate(ncode)
elif genre == 'sf':
synopsis = sf_s.generate(ncode)
score = rouge.get_scores(wakati(synopsis), wakati(ref),
False)[0]['rouge-1']['r']
file.write(ncode + '\n')
file.write(genre + '\n')
file.write('score: ' + str(score) + '\n')
file.write(ref + '\n\n')
file.write(synopsis + '\n\n\n')
file.close()
import os
import joblib
import numpy as np
from gensim.models import word2vec
import data_supplier
from data_supplier.active_ncodes_supplier import ncodes_train_test_split
from util.paths import LSTM_TRAINED_MODEL_DIR_PATH
from util.corpus_accessor import CorpusAccessor
from util.paths import WORD_EMBEDDING_MODEL_PATH
from util import text_processor
data_accessor = CorpusAccessor()
class LSTMVectorSupplier:
def __init__(self,
genre='general',
importance='cos_sim',
use_data_of_position_of_sentence=False,
use_data_of_is_serif=False,
use_data_of_is_include_person=False,
use_data_of_sentence_length=False):
"""
このクラスの初期化時に使用する素性を指定する
:param word_embedding_avg_vector: bool
単語の分散表現ベクトルの平均ベクトル
:param position_of_sentence: bool
文の出現位置
:param is_serif: bool
セリフか否か
