def create_model(args, maxlen, vocab, wordsref):
def ortho_reg(weight_matrix):
### orthogonal regularization for aspect embedding matrix ###
wn_ = weight_matrix / K.cast(K.epsilon() + K.sqrt(K.sum(K.square(WeightedSum), axis=-1, keepdims=True)), K.floatx())
reg = K.sum(K.square(K.dot(w_n, K.transpose(w_n))-K.eye(w_n.shape[0])))
return args.ortho_reg*reg
vocab_size = len(vocab)
################## Inputs #################
sentence_input = Input(shape=(maxlen,), dtype='int32', name='sentence_input')
neg_input = Input(shape=(args.neg_size, maxlen), dtype='int32', name='neg_input')
word_emb = Embedding(vocab_size, args.emb_dim, mask_zero=True, name='word_emb')
######### Compute sentence representation ########
e_w = word_emb(sentence_input)
e_wordsref = word_emb(wordsref)
y_s = Average()(e_w)
att_words_ref = WordsRef(name='att_words_ref')([e_wordsref, e_w, y_s])
z_ss = WeightedSum()([e_w, att_words_ref])
######## Compute representations of negative instances #######
e_neg = word_emb(neg_input)
z_n = Average()(e_neg)
def create_model(args, maxlen, vocab):
def ortho_reg(weight_matrix):
### orthogonal regularization for aspect embedding matrix ###
w_n = weight_matrix / K.cast(K.epsilon() + K.sqrt(K.sum(K.square(weight_matrix), axis=-1, keepdims=True)), K.floatx())
reg = K.sum(K.square(K.dot(w_n, K.transpose(w_n)) - K.eye((w_n.shape[0]).eval())))
return args.ortho_reg*reg
# 词汇表大小
vocab_size = len(vocab)
##### Inputs #####
# 正例的形状：batch_size * dim, 每个元素是在词汇表中的索引值, 每个句子有多少个词就有多少索引值
# 负例的形状：batch_size * args.neg_size * dim, ditto
# 得到w
sentence_input = Input(batch_shape=(None, maxlen), dtype='int32', name='sentence_input')
neg_input = Input(batch_shape=(None, args.neg_size, maxlen), dtype='int32', name='neg_input')
##### Construct word embedding layer #####
# 嵌入层将正整数（下标）转换为具有固定大小的向量，如[[4],[20]]->[[0.25,0.1],[0.6,-0.2]]
# keras.layers.embeddings.Embedding(input_dim, output_dim, embeddings_initializer='uniform', embeddings_regularizer=None, activity_regularizer=None, embeddings_constraint=None, mask_zero=False, input_length=None)
word_emb = Embedding(vocab_size, args.emb_dim, mask_zero=True, name='word_emb')
##### Compute sentence representation #####
# 计算句子嵌入，这里设计到keras的很多细节，日后还需要深入学习
e_w = word_emb(sentence_input)
y_s = Average()(e_w)
att_weights = Attention(name='att_weights')([e_w, y_s])
z_s = WeightedSum()([e_w, att_weights])
##### Compute representations of negative instances #####
# 计算负例的z_n
e_neg = word_emb(neg_input)
z_n = Average()(e_neg)
##### Reconstruction #####
# 重构过程
p_t = Dense(args.aspect_size)(z_s)
p_t = Activation('softmax', name='p_t')(p_t)
r_s = WeightedAspectEmb(args.aspect_size, args.emb_dim, name='aspect_emb',
W_regularizer=ortho_reg)(p_t)
##### Loss #####
# 损失函数
loss = MaxMargin(name='max_margin')([z_s, z_n, r_s])
model = Model(input=[sentence_input, neg_input], output=loss)
### Word embedding and aspect embedding initialization ######
# 如果定义了emb_path, 就用文件中的数值初始化E矩阵, T使用K-means初始化
if args.emb_path:
from w2vEmbReader import W2VEmbReader as EmbReader
emb_reader = EmbReader(args.emb_path, emb_dim=args.emb_dim)
logger.info('Initializing word embedding matrix')
model.get_layer('word_emb').set_weights(emb_reader.get_emb_matrix_given_vocab(vocab, model.get_layer('word_emb').get_weights()))
logger.info('Initializing aspect embedding matrix as centroid of kmean clusters')
model.get_layer('aspect_emb').W.set_value(emb_reader.get_aspect_matrix(args.aspect_size))
return model
def create_model(args, maxlen, vocab):
def ortho_reg(weight_matrix):
### orthogonal regularization for aspect embedding matrix ###
w_n = weight_matrix / K.cast(K.epsilon() + K.sqrt(K.sum(K.square(weight_matrix), axis=-1, keepdims=True)), K.floatx())
reg = K.sum(K.square(K.dot(w_n, K.transpose(w_n)) - K.eye(w_n.shape[0].value)))
return args.ortho_reg*reg
vocab_size = len(vocab)
##### Inputs #####
sentence_input = Input(shape=(maxlen,), dtype='int32', name='sentence_input')
neg_input = Input(shape=(args.neg_size, maxlen), dtype='int32', name='neg_input')
##### Construct word embedding layer #####
word_emb = Embedding(vocab_size, args.emb_dim, mask_zero=True, name='word_emb')
##### Compute sentence representation #####
e_w = word_emb(sentence_input)
#y_s = LSTM(args.emb_dim, return_sequences=False)(e_w)
y_s = Average()(e_w)
att_weights = Attention(name='att_weights')([e_w, y_s])
z_s = WeightedSum()([e_w, att_weights])
##### Compute representations of negative instances #####
e_neg = word_emb(neg_input)
#z_n = TimeDistributed(LSTM(args.emb_dim, return_sequences=False))(e_neg)
z_n = Average()(e_neg)
##### Reconstruction #####
p_t = Dense(args.aspect_size)(z_s)
p_t = Activation('softmax', name='p_t')(p_t)
r_s = WeightedAspectEmb(args.aspect_size, args.emb_dim, name='aspect_emb',
W_regularizer=ortho_reg)(p_t)
##### Loss #####
loss = MaxMargin(name='max_margin')([z_s, z_n, r_s])
model = Model(inputs=[sentence_input, neg_input], outputs=loss)
### Word embedding and aspect embedding initialization ######
if args.emb_path:
from w2vEmbReader import W2VEmbReader as EmbReader
emb_reader = EmbReader(args.emb_path, emb_dim=args.emb_dim)
logger.info('Initializing word embedding matrix')
# model.get_layer('word_emb').set_weights(emb_reader.get_emb_matrix_given_vocab(vocab, model.get_layer('word_emb').get_weights()[0]))
K.set_value(
model.get_layer('word_emb').embeddings,
emb_reader.get_emb_matrix_given_vocab(vocab, K.get_value(model.get_layer('word_emb').embeddings)))
logger.info('Initializing aspect embedding matrix as centroid of kmean clusters')
K.set_value(
model.get_layer('aspect_emb').W,
emb_reader.get_aspect_matrix(args.aspect_size))
return model
###############################################################################################################################
## Building model
sentence_input = Input(shape=(args.kstep, node_size),
dtype='float32',
name='sentence_inputt')
neg_input = Input(shape=(args.neg_size, args.kstep, node_size),
dtype='float32',
name='neg_inputt')
predict = Input(shape=(17, ), dtype='int32', name='predictt')
e_w = sentence_input
y_s = Average()(sentence_input)
att_weights = Attention(name='att_weights')([e_w, y_s])
z_s = WeightedSum()([e_w, att_weights])
##### Compute representations of negative instances #####
e_neg = neg_input
z_n = Average()(e_neg)
##### Reconstruction2 #####
dense1 = Dense(512, activation='relu')(z_s)
#dense1 = noise.GaussianNoise(0.09)(dense1)
p_t = Dense(128, activation='relu')(dense1)
dense2 = Dense(512, activation='relu')(p_t)
new_p_t = Activation('softmax', name='p_t')(dense2)
r_s = Dense(node_size, init='uniform')(new_p_t)
##### Loss1 #####
def create_model(args, maxlen, vocab):
def ortho_reg(weight_matrix):
### orthogonal regularization for aspect embedding matrix ###
w_n = K.l2_normalize(weight_matrix, axis=-1)
reg = K.sum(
K.square(
K.dot(w_n, K.transpose(w_n)) -
K.eye(w_n.get_shape().as_list()[0])))
return args.ortho_reg * reg
vocab_size = len(vocab)
if args.emb_name:
from w2vEmbReader import W2VEmbReader as EmbReader
emb_reader = EmbReader(
os.path.join("..", "preprocessed_data", args.domain),
args.emb_name)
aspect_matrix = emb_reader.get_aspect_matrix(args.aspect_size)
args.aspect_size = emb_reader.aspect_size
args.emb_dim = emb_reader.emb_dim
##### Inputs #####
sentence_input = Input(shape=(maxlen, ),
dtype='int32',
name='sentence_input')
neg_input = Input(shape=(args.neg_size, maxlen),
dtype='int32',
name='neg_input')
##### Construct word embedding layer #####
word_emb = Embedding(vocab_size,
args.emb_dim,
mask_zero=True,
name='word_emb')
#, embeddings_constraint=MaxNorm(10))
##### Compute sentence representation #####
e_w = word_emb(sentence_input)
y_s = Average()(e_w)
att_weights = Attention(name='att_weights')([e_w, y_s])
#W_constraint=MaxNorm(10),
#b_constraint=MaxNorm(10))([e_w, y_s])
z_s = WeightedSum()([e_w, att_weights])
##### Compute representations of negative instances #####
e_neg = word_emb(neg_input)
z_n = Average()(e_neg)
##### Reconstruction #####
p_t = Dense(args.aspect_size)(z_s)
p_t = Activation('softmax', name='p_t')(p_t)
r_s = WeightedAspectEmb(
args.aspect_size,
args.emb_dim,
name='aspect_emb',
#W_constraint=MaxNorm(10),
W_regularizer=ortho_reg)(p_t)
##### Loss #####
loss = MaxMargin(name='max_margin')([z_s, z_n, r_s])
model = Model(inputs=[sentence_input, neg_input], outputs=[loss])
### Word embedding and aspect embedding initialization ######
if args.emb_name:
from w2vEmbReader import W2VEmbReader as EmbReader
logger.info('Initializing word embedding matrix')
embs = model.get_layer('word_emb').embeddings
K.set_value(
embs,
emb_reader.get_emb_matrix_given_vocab(vocab, K.get_value(embs)))
logger.info(
'Initializing aspect embedding matrix as centroid of kmean clusters'
)
K.set_value(model.get_layer('aspect_emb').W, aspect_matrix)
return model
def create_model(args, kstep, node_size):
def ortho_reg(weight_matrix):
### orthogonal regularization for aspect embedding matrix ###
w_n = weight_matrix / K.cast(K.epsilon() + K.sqrt(K.sum(K.square(weight_matrix), axis=-1, keepdims=True)), K.floatx())
reg = K.sum(K.square(K.dot(w_n, K.transpose(w_n)) - K.eye(w_n.shape[0].eval())))
return args.ortho_reg*reg
##### Inputs #####
sentence_input = Input(shape=(kstep, node_size), dtype='float32', name='sentence_input')
neg_input = Input(shape=(args.neg_size, kstep, node_size), dtype='float32', name='neg_input')
print("sentence_input.ndim", sentence_input.ndim)
print("neg_input.ndim", neg_input.ndim)
e_w = sentence_input
y_s = Average()(sentence_input)
print(y_s.ndim)
print(e_w.ndim)
print(K.int_shape(e_w), K.int_shape(y_s))
att_weights = Attention(name='att_weights')([e_w, y_s])
z_s = WeightedSum()([e_w, att_weights])
print("z_s----------- %d", (z_s.ndim))
##### Compute representations of negative instances #####
#e_neg = word_emb(neg_input)
e_neg = neg_input
z_n = Average()(e_neg)
print("e_neg.ndim")
print(e_neg.ndim)
print("z_n.ndim")
print(z_n.ndim)
##### Reconstruction #####
p_t = Dense(args.aspect_size)(z_s)
p_t = Activation('softmax', name='p_t')(p_t)
r_s = WeightedAspectEmb(args.aspect_size, 2405, name='aspect_emb',
W_regularizer=ortho_reg)(p_t)
##### Loss #####
print("losssssssssssssssssssssssssssssssssssssssssssssssssssssssssss")
print(K.int_shape(z_s), K.int_shape(r_s))
loss = MaxMargin(name='max_margin')([z_s, z_n, r_s])
model = Model(input=[sentence_input, neg_input], output=loss)
### Word embedding and aspect embedding initialization ######
if args.emb_path:
from w2vEmbReader import W2VEmbReader as EmbReader
emb_reader = EmbReader(args.emb_path, emb_dim=args.emb_dim)
logger.info('Initializing word embedding matrix')
model.get_layer('word_emb').W.set_value(emb_reader.get_emb_matrix_given_vocab(vocab, model.get_layer('word_emb').W.get_value()))
logger.info('Initializing aspect embedding matrix as centroid of kmean clusters')
model.get_layer('aspect_emb').W.set_value(emb_reader.get_aspect_matrix(args.aspect_size))
return model
def create_model(args, maxlen, vocab):
def ortho_reg(weight_matrix):
### orthogonal regularization for aspect embedding matrix ###
w_n = K.l2_normalize(weight_matrix, axis=-1) # K表示调用该函数的当前layer
reg = K.sum(
K.square(
K.dot(w_n, K.transpose(w_n)) -
K.eye(w_n.shape[0].value))) # 自身矩阵的内积的平方根-自身特征值 = 越小, 越说明分量为0
return args.ortho_reg * reg # 这东西越小越好, 因为能保证各个特征分的越开
vocab_size = len(vocab)
if args.emb_name: # 获取已经保存的embedding???
from w2vEmbReader import W2VEmbReader as EmbReader
emb_reader = EmbReader(
os.path.join(
"/content/drive/My Drive/Attention-Based-Aspect-Extraction-master",
"preprocessed_data", args.domain), args.emb_name)
aspect_matrix = emb_reader.get_aspect_matrix(args.aspect_size)
args.aspect_size = emb_reader.aspect_size
args.emb_dim = emb_reader.emb_dim
##### Inputs #####
sentence_input = Input(shape=(maxlen, ),
dtype='int32',
name='sentence_input')
neg_input = Input(shape=(args.neg_size, maxlen),
dtype='int32',
name='neg_input')
##### Construct word embedding layer #####
word_emb = Embedding(vocab_size,
args.emb_dim,
mask_zero=True,
name='word_emb',
embeddings_constraint=MaxNorm(10))
##### Compute sentence representation ##### pre-processing 根据attention组合句子
e_w = word_emb(sentence_input) # 将input转换为embedding
y_s = Average()(e_w) # 默认求平均 layer
att_weights = Attention(name='att_weights',
W_constraint=MaxNorm(10),
b_constraint=MaxNorm(10))([e_w,
y_s]) # attention layer
z_s = WeightedSum()([e_w, att_weights]) # encoding layer
##### Compute representations of negative instances ##### 增加准确性的tricks
e_neg = word_emb(neg_input)
z_n = Average()(e_neg)
##### Reconstruction ##### 构建dense层, 希望能够decoding attention sentences的特征
p_t = Dense(args.aspect_size)(z_s)
p_t = Activation('softmax',
name='p_t')(p_t) # softmax一下, nodes数量不改变, 数值被soft了一下
r_s = WeightedAspectEmb(args.aspect_size,
args.emb_dim,
name='aspect_emb',
W_constraint=MaxNorm(10),
W_regularizer=ortho_reg)(
p_t) # 标准化0-10区间, 且正则项为自定义的ortho_reg
##### Loss #####
loss = MaxMargin(name='max_margin')([z_s, z_n,
r_s]) # 自定义loss function??? 这是在做啥???
model = Model(inputs=[sentence_input, neg_input],
outputs=[loss]) # negative input是需要自己分开数据集的吗??
### Word embedding and aspect embedding initialization ######
if args.emb_name:
from w2vEmbReader import W2VEmbReader as EmbReader
logger.info('Initializing word embedding matrix')
embs = model.get_layer('word_emb').embeddings
K.set_value(
embs,
emb_reader.get_emb_matrix_given_vocab(vocab, K.get_value(embs)))
logger.info(
'Initializing aspect embedding matrix as centroid of kmean clusters'
) # 为何初始化要用到kmeans
K.set_value(model.get_layer('aspect_emb').W, aspect_matrix) # r-s
return model
def create_model(args, maxlen, vocab):
def ortho_reg(weight_matrix):
### orthogonal regularization for aspect embedding matrix ###
w_n = K.l2_normalize(weight_matrix, axis=-1)
reg = K.sum(
K.square(K.dot(w_n, K.transpose(w_n)) - K.eye(w_n.shape[0].value)))
return args.ortho_reg * reg
vocab_size = len(vocab)
if args.emb_name:
if args.emb_technique == "w2v":
logger.info("Load {} glove embedding for {}".format(
args.lang, config.word_emb_training_type))
if args.lang == 'en':
emb_reader = EmbReader(
config.emb_dir_en["w2v"].format(
config.word_emb_training_type), args.emb_name)
elif args.lang == 'de':
#emb_reader = EmbReader(config.emb_dir_de["w2v"].format(config.word_emb_training_type),
#args.emb_name)
emb_reader = FineTuneEmbed_cca(
'../preprocessed_data/german/w2v/fine_tuned', 'w2v_emb',
'../preprocessed_data/german/w2v/full_trained',
'w2v_embedding_300')
elif args.emb_technique == 'fasttext':
if args.lang == 'en':
#emb_reader = FastTextEmbReader(config.emb_dir_en["fasttext"].format(config.word_emb_training_type),
#args.emb_name, config.fine_tuned_enabled)
emb_reader = FineTuneEmbed_ortho_procrustes(
'../preprocessed_data/fasttext/fine_tuned',
'fasttext_pre_trained',
'../preprocessed_data/fasttext/full_trained',
'w2v_embedding_skipgram_300')
elif args.lang == 'de':
emb_reader = FastTextEmbReader(
config.emb_dir_de["fasttext"].format(
config.word_emb_training_type), args.emb_name,
config.fine_tuned_enabled)
#emb_reader = FineTuneEmbed_ortho_procrustes('../preprocessed_data/fasttext/fine_tuned','fasttext_pre_trained','../preprocessed_data/fasttext/full_trained', 'w2v_embedding_skipgram_300')
elif args.emb_technique == "glove":
if args.lang == 'de':
logger.info('Load german glove embedding')
emb_reader = GloveEmbedding(config.emb_dir_de["glove"],
args.emb_name)
else:
logger.info("Load en glove embedding for {}".format(
config.word_emb_training_type))
emb_reader = GloveEmbedding(
config.emb_dir_en["glove"].format(
config.word_emb_training_type), args.emb_name)
elif args.emb_technique == "MUSE_supervised":
emb_reader = MUSEEmbedding(config.emb_dir_biling['supervised'],
args.emb_name)
elif args.emb_technique == "MUSE_unsupervised":
emb_reader = MUSEEmbedding(config.emb_dir_biling['unsupervised'],
args.emb_name)
aspect_matrix = emb_reader.get_aspect_matrix(args.aspect_size)
args.aspect_size = emb_reader.aspect_size
args.emb_dim = emb_reader.emb_dim
##### Inputs #####
sentence_input = Input(shape=(maxlen, ),
dtype='int32',
name='sentence_input')
neg_input = Input(shape=(args.neg_size, maxlen),
dtype='int32',
name='neg_input')
##### Construct word embedding layer #####
word_emb = Embedding(vocab_size,
args.emb_dim,
mask_zero=True,
name='word_emb',
embeddings_constraint=MaxNorm(10))
##### Compute sentence representation #####
e_w = word_emb(sentence_input)
y_s = Average()(e_w)
att_weights = Attention(name='att_weights',
W_constraint=MaxNorm(10),
b_constraint=MaxNorm(10))([e_w, y_s])
z_s = WeightedSum()([e_w, att_weights])
##### Compute representations of negative instances #####
e_neg = word_emb(neg_input)
z_n = Average()(e_neg)
##### Reconstruction #####
p_t = Dense(args.aspect_size)(z_s)
p_t = Activation('softmax', name='p_t')(p_t)
r_s = WeightedAspectEmb(args.aspect_size,
args.emb_dim,
name='aspect_emb',
W_constraint=MaxNorm(10),
W_regularizer=ortho_reg)(p_t)
##### Loss #####
loss = MaxMargin(name='max_margin')([z_s, z_n, r_s])
model = Model(inputs=[sentence_input, neg_input], outputs=[loss])
### Word embedding and aspect embedding initialization ######
if args.emb_name:
logger.info('Initializing word embedding matrix')
embs = model.get_layer('word_emb').embeddings
K.set_value(
embs,
emb_reader.get_emb_matrix_given_vocab(vocab, K.get_value(embs)))
logger.info(
'Initializing aspect embedding matrix as centroid of kmean clusters'
)
K.set_value(model.get_layer('aspect_emb').W, aspect_matrix)
return model
def create_model(args, vocab, num_outputs, overal_maxlen, maxlen_aspect):
###############################################################################################################################
## Recurrence unit type
#
if args.recurrent_unit == 'lstm':
from keras.layers.recurrent import LSTM as RNN
elif args.recurrent_unit == 'gru':
from keras.layers.recurrent import GRU as RNN
elif args.recurrent_unit == 'simple':
from keras.layers.recurrent import SimpleRNN as RNN
###############################################################################################################################
## Create Model
#
dropout = args.dropout_W
recurrent_dropout = args.dropout_U
vocab_size = len(vocab)
logger.info('Building a LSTM attention model to predict term/aspect sentiment')
print '\n\n'
##### Inputs #####
sentence_input = Input(shape=(overal_maxlen,), dtype='int32', name='sentence_input')
aspect_input = Input(shape=(maxlen_aspect,), dtype='int32', name='aspect_input')
pretrain_input = Input(shape=(None,), dtype='int32', name='pretrain_input')
##### construct word embedding layer #####
word_emb = Embedding(vocab_size, args.emb_dim, mask_zero=True, name='word_emb')
### represent aspect as averaged word embedding ###
print 'use average term embs as aspect embedding'
aspect_term_embs = word_emb(aspect_input)
aspect_embs = Average(mask_zero=True, name='aspect_emb')(aspect_term_embs)
### sentence representation ###
sentence_output = word_emb(sentence_input)
pretrain_output = word_emb(pretrain_input)
print 'use a rnn layer'
rnn = RNN(args.rnn_dim, return_sequences=True, dropout=dropout, recurrent_dropout=recurrent_dropout, name='lstm')
sentence_output = rnn(sentence_output)
pretrain_output = rnn(pretrain_output)
print 'use content attention to get term weights'
att_weights = Attention(name='att_weights')([sentence_output, aspect_embs])
sentence_output = WeightedSum()([sentence_output, att_weights])
pretrain_output = Average(mask_zero=True)(pretrain_output)
if args.dropout_prob > 0:
print 'use dropout layer'
sentence_output = Dropout(args.dropout_prob)(sentence_output)
pretrain_output = Dropout(args.dropout_prob)(pretrain_output)
sentence_output = Dense(num_outputs, name='dense_1')(sentence_output)
pretrain_output = Dense(num_outputs, name='dense_2')(pretrain_output)
aspect_probs = Activation('softmax', name='aspect_model')(sentence_output)
doc_probs = Activation('softmax', name='pretrain_model')(pretrain_output)
model = Model(inputs=[sentence_input, aspect_input, pretrain_input], outputs=[aspect_probs, doc_probs])
logger.info(' Done')
###############################################################################################################################
## Initialize embeddings if requested
#
if args.is_pretrain:
import pickle
print 'Set embedding, lstm, and dense weights from pre-trained models'
if args.domain == 'lt':
f_1 = open('../pretrained_weights/lstm_weights_lt%.1f.pkl'%(args.percetage), 'rb')
f_2 = open('../pretrained_weights/dense_weights_lt%.1f.pkl'%(args.percetage), 'rb')
else:
f_1 = open('../pretrained_weights/lstm_weights_res%.1f.pkl'%(args.percetage), 'rb')
f_2 = open('../pretrained_weights/dense_weights_res%.1f.pkl'%(args.percetage), 'rb')
lstm_weights = pickle.load(f_1)
dense_weights = pickle.load(f_2)
model.get_layer('lstm').set_weights(lstm_weights)
model.get_layer('dense_1').set_weights(dense_weights)
model.get_layer('dense_2').set_weights(dense_weights)
from w2vEmbReader import W2VEmbReader as EmbReader
logger.info('Initializing lookup table')
emb_path = '../glove/%s.txt'%(args.domain)
emb_reader = EmbReader(args, emb_path)
model.get_layer('word_emb').set_weights(emb_reader.get_emb_matrix_given_vocab(vocab, model.get_layer('word_emb').get_weights()))
logger.info(' Done')
return model
def create_model(args, vocab, nb_class, overall_maxlen, doc_maxlen_1,
doc_maxlen_2):
# Funtion that initializes word embeddings
def init_emb(emb_matrix, vocab, emb_file_gen, emb_file_domain):
print 'Loading pretrained general word embeddings and domain word embeddings ...'
counter_gen = 0.
pretrained_emb = open(emb_file_gen)
for line in pretrained_emb:
tokens = line.split()
if len(tokens) != 301:
continue
word = tokens[0]
vec = tokens[1:]
try:
emb_matrix[0][vocab[word]][:300] = vec
counter_gen += 1
except KeyError:
pass
if args.use_domain_emb:
counter_domain = 0.
pretrained_emb = open(emb_file_domain)
for line in pretrained_emb:
tokens = line.split()
if len(tokens) != 101:
continue
word = tokens[0]
vec = tokens[1:]
try:
emb_matrix[0][vocab[word]][300:] = vec
counter_domain += 1
except KeyError:
pass
pretrained_emb.close()
logger.info(
'%i/%i word vectors initialized by general embeddings (hit rate: %.2f%%)'
% (counter_gen, len(vocab), 100 * counter_gen / len(vocab)))
if args.use_domain_emb:
logger.info(
'%i/%i word vectors initialized by domain embeddings (hit rate: %.2f%%)'
% (counter_domain, len(vocab),
100 * counter_domain / len(vocab)))
return emb_matrix
# Build model
logger.info('Building model ...')
print 'Building model ...'
print '\n\n'
vocab_size = len(vocab)
###################################
# Inputs
###################################
print 'Input layer'
# sequence of token indices for aspect-level data
sentence_input = Input(shape=(overall_maxlen, ),
dtype='int32',
name='sentence_input')
# gold opinion label for aspect-level data.
op_label_input = Input(shape=(overall_maxlen, 3),
dtype=K.floatx(),
name='op_label_input')
# probability of sending gold opinion labels at opinion transmission step
p_gold_op = Input(shape=(overall_maxlen, ),
dtype=K.floatx(),
name='p_gold_op')
if args.use_doc:
# doc_input_1 denotes the data for sentiment classification
# doc_input_2 denotes the data for domain classification
doc_input_1 = Input(shape=(doc_maxlen_1, ),
dtype='int32',
name='doc_input_1')
doc_input_2 = Input(shape=(doc_maxlen_2, ),
dtype='int32',
name='doc_input_2')
#########################################
# Shared word embedding layer
#########################################
print 'Word embedding layer'
word_emb = Embedding(vocab_size,
args.emb_dim,
mask_zero=True,
name='word_emb')
# aspect-level inputs
word_embeddings = word_emb(sentence_input)
sentence_output = word_embeddings
# doc-level inputs
if args.use_doc:
doc_output_1 = word_emb(doc_input_1)
# we only use general embedding for domain classification
doc_output_2 = word_emb(doc_input_2)
if args.use_domain_emb:
# mask out the domain embeddings
doc_output_2 = Remove_domain_emb()(doc_output_2)
######################################
# Shared CNN layers
######################################
for i in xrange(args.shared_layers):
print 'Shared CNN layer %s' % i
sentence_output = Dropout(args.dropout_prob)(sentence_output)
if args.use_doc:
doc_output_1 = Dropout(args.dropout_prob)(doc_output_1)
doc_output_2 = Dropout(args.dropout_prob)(doc_output_2)
if i == 0:
conv_1 = Conv1DWithMasking(filters=args.cnn_dim/2, kernel_size=3, \
activation='relu', padding='same', kernel_initializer=my_init, name='cnn_0_1')
conv_2 = Conv1DWithMasking(filters=args.cnn_dim/2, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='cnn_0_2')
sentence_output_1 = conv_1(sentence_output)
sentence_output_2 = conv_2(sentence_output)
sentence_output = Concatenate()(
[sentence_output_1, sentence_output_2])
if args.use_doc:
doc_output_1_1 = conv_1(doc_output_1)
doc_output_1_2 = conv_2(doc_output_1)
doc_output_1 = Concatenate()([doc_output_1_1, doc_output_1_2])
doc_output_2_1 = conv_1(doc_output_2)
doc_output_2_2 = conv_2(doc_output_2)
doc_output_2 = Concatenate()([doc_output_2_1, doc_output_2_2])
else:
conv = Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='cnn_%s'%i)
sentence_output = conv(sentence_output)
if args.use_doc:
doc_output_1 = conv(doc_output_1)
doc_output_2 = conv(doc_output_2)
word_embeddings = Concatenate()([word_embeddings, sentence_output])
init_shared_features = sentence_output
#######################################
# Define task-specific layers
#######################################
# AE specific layers
aspect_cnn = Sequential()
for a in xrange(args.aspect_layers):
print 'Aspect extraction layer %s' % a
aspect_cnn.add(Dropout(args.dropout_prob))
aspect_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='aspect_cnn_%s'%a))
aspect_dense = Dense(nb_class, activation='softmax', name='aspect_dense')
# AS specific layers
sentiment_cnn = Sequential()
for b in xrange(args.senti_layers):
print 'Sentiment classification layer %s' % b
sentiment_cnn.add(Dropout(args.dropout_prob))
sentiment_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='sentiment_cnn_%s'%b))
sentiment_att = Self_attention(args.use_opinion, name='sentiment_att')
sentiment_dense = Dense(3, activation='softmax', name='sentiment_dense')
if args.use_doc:
# DS specific layers
doc_senti_cnn = Sequential()
for c in xrange(args.doc_senti_layers):
print 'Document-level sentiment layers %s' % c
doc_senti_cnn.add(Dropout(args.dropout_prob))
doc_senti_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='doc_sentiment_cnn_%s'%c))
doc_senti_att = Attention(name='doc_senti_att')
doc_senti_dense = Dense(3, name='doc_senti_dense')
# The reason not to use the default softmax is that it reports errors when input_dims=2 due to
# compatibility issues between the tf and keras versions used.
softmax = Lambda(lambda x: K.tf.nn.softmax(x),
name='doc_senti_softmax')
# DD specific layers
doc_domain_cnn = Sequential()
for d in xrange(args.doc_domain_layers):
print 'Document-level domain layers %s' % d
doc_domain_cnn.add(Dropout(args.dropout_prob))
doc_domain_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='doc_domain_cnn_%s'%d))
doc_domain_att = Attention(name='doc_domain_att')
doc_domain_dense = Dense(1,
activation='sigmoid',
name='doc_domain_dense')
# re-encoding layer
enc = Dense(args.cnn_dim, activation='relu', name='enc')
####################################################
# aspect-level operations involving message passing
####################################################
for i in xrange(args.interactions + 1):
print 'Interaction number ', i
aspect_output = sentence_output
sentiment_output = sentence_output
# note that the aspet-level data will also go through the doc-level models
doc_senti_output = sentence_output
doc_domain_output = sentence_output
### AE ###
if args.aspect_layers > 0:
aspect_output = aspect_cnn(aspect_output)
# concate word embeddings and task-specific output for prediction
aspect_output = Concatenate()([word_embeddings, aspect_output])
aspect_output = Dropout(args.dropout_prob)(aspect_output)
aspect_probs = aspect_dense(aspect_output)
### AS ###
if args.senti_layers > 0:
sentiment_output = sentiment_cnn(sentiment_output)
sentiment_output = sentiment_att(
[sentiment_output, op_label_input, aspect_probs, p_gold_op])
sentiment_output = Concatenate()(
[init_shared_features, sentiment_output])
sentiment_output = Dropout(args.dropout_prob)(sentiment_output)
sentiment_probs = sentiment_dense(sentiment_output)
if args.use_doc:
### DS ###
if args.doc_senti_layers > 0:
doc_senti_output = doc_senti_cnn(doc_senti_output)
# output attention weights with two activation functions
senti_att_weights_softmax, senti_att_weights_sigmoid = doc_senti_att(
doc_senti_output)
# reshape the sigmoid attention weights, will be used in message passing
senti_weights = Lambda(lambda x: K.expand_dims(x, axis=-1))(
senti_att_weights_sigmoid)
doc_senti_output = WeightedSum()(
[doc_senti_output, senti_att_weights_softmax])
doc_senti_output = Dropout(args.dropout_prob)(doc_senti_output)
doc_senti_output = doc_senti_dense(doc_senti_output)
doc_senti_probs = softmax(doc_senti_output)
# reshape the doc-level sentiment predictions, will be used in message passing
doc_senti_probs = Lambda(lambda x: K.expand_dims(x, axis=-2))(
doc_senti_probs)
doc_senti_probs = Lambda(lambda x: K.repeat_elements(
x, overall_maxlen, axis=1))(doc_senti_probs)
### DD ###
if args.doc_domain_layers > 0:
doc_domain_output = doc_domain_cnn(doc_domain_output)
domain_att_weights_softmax, domain_att_weights_sigmoid = doc_domain_att(
doc_domain_output)
domain_weights = Lambda(lambda x: K.expand_dims(x, axis=-1))(
domain_att_weights_sigmoid)
doc_domain_output = WeightedSum()(
[doc_domain_output, domain_att_weights_softmax])
doc_domain_output = Dropout(args.dropout_prob)(doc_domain_output)
doc_domain_probs = doc_domain_dense(doc_domain_output)
# update sentence_output for the next iteration
sentence_output = Concatenate()([
sentence_output, aspect_probs, sentiment_probs,
doc_senti_probs, senti_weights, domain_weights
])
else:
# update sentence_output for the next iteration
sentence_output = Concatenate()(
[sentence_output, aspect_probs, sentiment_probs])
sentence_output = enc(sentence_output)
aspect_model = Model(inputs=[sentence_input, op_label_input, p_gold_op],
outputs=[aspect_probs, sentiment_probs])
####################################################
# doc-level operations without message passing
####################################################
if args.use_doc:
if args.doc_senti_layers > 0:
doc_output_1 = doc_senti_cnn(doc_output_1)
att_1, _ = doc_senti_att(doc_output_1)
doc_output_1 = WeightedSum()([doc_output_1, att_1])
doc_output_1 = Dropout(args.dropout_prob)(doc_output_1)
doc_output_1 = doc_senti_dense(doc_output_1)
doc_prob_1 = softmax(doc_output_1)
if args.doc_domain_layers > 0:
doc_output_2 = doc_domain_cnn(doc_output_2)
att_2, _ = doc_domain_att(doc_output_2)
doc_output_2 = WeightedSum()([doc_output_2, att_2])
doc_output_2 = Dropout(args.dropout_prob)(doc_output_2)
doc_prob_2 = doc_domain_dense(doc_output_2)
doc_model = Model(inputs=[doc_input_1, doc_input_2],
outputs=[doc_prob_1, doc_prob_2])
else:
doc_model = None
####################################################
# initialize word embeddings
####################################################
logger.info('Initializing lookup table')
# Load pre-trained word vectors.
# To save the loading time, here we load from the extracted subsets of the original embeddings,
# which only contains the embeddings of words in the vocab.
if args.use_doc:
emb_path_gen = '../glove/%s_.txt' % (args.domain)
emb_path_domain = '../domain_specific_emb/%s_.txt' % (args.domain)
else:
emb_path_gen = '../glove/%s.txt' % (args.domain)
emb_path_domain = '../domain_specific_emb/%s.txt' % (args.domain)
# Load pre-trained word vectors from the orginal large files
# If you are loading from ssd, the process would only take 1-2 mins
# If you are loading from hhd, the process would take a few hours at first try,
# and would take 1-2 mins in subsequent repeating runs (due to cache performance).
# emb_path_gen = '../glove.840B.300d.txt'
# if args.domain == 'lt':
# emb_path_domain = '../laptop_emb.vec'
# else:
# emb_path_domain = '../restaurant_emb.vec'
aspect_model.get_layer('word_emb').set_weights(
init_emb(
aspect_model.get_layer('word_emb').get_weights(), vocab,
emb_path_gen, emb_path_domain))
logger.info(' Done')
return aspect_model, doc_model
def create_model(args, vocab, nb_class, overall_maxlen, doc_maxlen_1,
doc_maxlen_2, num_relations):
# Funtion that initializes word embeddings
def init_emb(emb_matrix, vocab, emb_file_gen, emb_file_domain):
print 'Loading pretrained general word embeddings and domain word embeddings ...'
counter_gen = 0.
pretrained_emb = open(emb_file_gen)
for line in pretrained_emb:
tokens = line.split()
if len(tokens) != 301:
continue
word = tokens[0]
vec = tokens[1:]
try:
emb_matrix[0][vocab[word]][:300] = vec
counter_gen += 1
except KeyError:
pass
if args.use_domain_emb:
counter_domain = 0.
pretrained_emb = open(emb_file_domain)
for line in pretrained_emb:
tokens = line.split()
if len(tokens) != 101:
continue
word = tokens[0]
vec = tokens[1:]
try:
emb_matrix[0][vocab[word]][300:] = vec
counter_domain += 1
except KeyError:
pass
pretrained_emb.close()
logger.info(
'%i/%i word vectors initialized by general embeddings (hit rate: %.2f%%)'
% (counter_gen, len(vocab), 100 * counter_gen / len(vocab)))
if args.use_domain_emb:
logger.info(
'%i/%i word vectors initialized by domain embeddings (hit rate: %.2f%%)'
% (counter_domain, len(vocab),
100 * counter_domain / len(vocab)))
return emb_matrix
# Build model
logger.info('Building model ...')
print 'Building model ...'
print '\n\n'
vocab_size = len(vocab)
###################################
# Inputs
###################################
print 'Input layer'
# sequence of token indices for aspect-level data
sentence_input = Input(shape=(overall_maxlen, ),
dtype='int32',
name='sentence_input')
A_in = [
Input(shape=(overall_maxlen, overall_maxlen), name='A_in%s' % i)
for i in range(num_relations)
]
# gold opinion label for aspect-level data.
op_label_input = Input(shape=(overall_maxlen, 3),
dtype=K.floatx(),
name='op_label_input')
# probability of sending gold opinion labels at opinion transmission step
p_gold_op = Input(shape=(overall_maxlen, ),
dtype=K.floatx(),
name='p_gold_op')
mask = K.not_equal(sentence_input, 0)
if args.use_doc:
# doc_input_1 denotes the data for sentiment classification
# doc_input_2 denotes the data for domain classification
doc_input_1 = Input(shape=(doc_maxlen_1, ),
dtype='int32',
name='doc_input_1')
doc_input_2 = Input(shape=(doc_maxlen_2, ),
dtype='int32',
name='doc_input_2')
if args.use_bert:
if args.bert_type == 'base':
hs = 768
bert_input = Input(shape=(overall_maxlen + 1, hs),
dtype=K.floatx(),
name='bert_input') # +1 denote +cls
#########################################
# Shared word embedding layer
#########################################
print 'Word embedding layer'
word_emb = Embedding(vocab_size,
args.emb_dim,
mask_zero=True,
name='word_emb')
# aspect-level inputs
word_embeddings = word_emb(sentence_input)
sentence_output = word_embeddings
# doc-level inputs
if args.use_doc:
doc_output_1 = word_emb(doc_input_1)
# we only use general embedding for domain classification
doc_output_2 = word_emb(doc_input_2)
if args.use_domain_emb:
# mask out the domain embeddings
doc_output_2 = Remove_domain_emb()(doc_output_2)
def slice(x, index):
return x[:, index, :]
def slice1(x, index):
return x[:, index:, :]
expand_dim = Lambda(lambda x: K.expand_dims(x, axis=1))
if args.use_bert:
#code.interact(local=locals())
bert_inp = Lambda(slice1, arguments={'index': 1})(bert_input)
bert_cls = Lambda(slice, arguments={'index': 0})(bert_input)
sentence_output = Concatenate()([sentence_output, bert_inp])
# if args.use_bert_cls:
#code.interact(local=locals())
#bert_cls = bert_input[:,0,:]
node_num = sentence_output.shape.as_list()[1]
bert_cls1 = expand_dim(bert_cls)
bert_cls = Lambda(lambda x: K.tile(x, [1, node_num, 1]))(bert_cls1)
if args.use_bert_cls == 0 and args.use_bert:
sentence_output = Concatenate()([sentence_output, bert_cls])
######################################
# Shared GCN + CNN layers
######################################
# iter_gcn = Sequential()
# iter_gcn.add(SpectralGraphConvolution(150, activation='relu',name='GCN'))
expand_dim = Lambda(lambda x: K.expand_dims(x, axis=1))
share_gcn_dense = Dense(300, activation='relu', name='share_gcn_dense')
for i in xrange(args.shared_layers):
print 'Shared GCN layer %s' % i
sentence_output = Dropout(args.dropout_prob)(sentence_output)
if args.use_doc:
doc_output_1 = Dropout(args.dropout_prob)(doc_output_1)
doc_output_2 = Dropout(args.dropout_prob)(doc_output_2)
if i == 0:
gcn_0 = SpectralGraphConvolution(args.gcn_dim,
args.relation_dim,
activation='relu',
name='GCN_0')
# conv_2 = SpectralGraphConvolution(args.cnn_dim, activation='relu',name='GCN_0_2')
#expand_dim = Lambda(lambda x: K.expand_dims(x, axis = 1))
sentence_output_0 = gcn_0([sentence_output] + A_in)
H = Dropout(args.dropout_prob)(sentence_output_0)
Global_graph = MyMeanPool(axis=1, smask=mask)(H)
node_num = H.shape.as_list()[1]
Gg = expand_dim(Global_graph)
GG = Lambda(lambda x: K.tile(x, [1, node_num, 1]))(Gg)
# code.interact(local=locals())
# HG = Concatenate()([H, GG])
# sentence_output = HG
if args.use_cnn:
conv_1 = Conv1DWithMasking(filters=args.cnn_dim, kernel_size=3, \
activation='relu', padding='same', kernel_initializer=my_init, name='CNN_0_1')
conv_2 = Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='CNN_0_2')
sentence_output_1 = conv_1(sentence_output)
sentence_output_2 = conv_2(sentence_output)
if args.use_meanpool:
sentence_output = Concatenate()(
[sentence_output_1, sentence_output_2, GG, H])
else:
sentence_output = Concatenate()(
[sentence_output_1, sentence_output_2, H])
else:
if args.use_meanpool:
sentence_output = Concatenate()([GG, H])
else:
sentence_output = H
if args.use_bert_cls == 1 and args.use_bert:
sentence_output = Concatenate()([sentence_output, bert_cls])
if args.use_doc:
doc_output_1_1 = conv_1(doc_output_1)
doc_output_1_2 = conv_2(doc_output_1)
doc_output_1 = Concatenate()([doc_output_1_1, doc_output_1_2])
doc_output_2_1 = conv_1(doc_output_2)
doc_output_2_2 = conv_2(doc_output_2)
doc_output_2 = Concatenate()([doc_output_2_1, doc_output_2_2])
if args.shared_layers == 1:
sentence_output = share_gcn_dense(sentence_output)
else:
# conv = Conv1DWithMasking(filters=args.cnn_dim/3, kernel_size=5, \
# activation='relu', padding='same', kernel_initializer=my_init, name='CNN_%s'%i)
gcn = SpectralGraphConvolution(args.gcn_dim,
args.relation_dim,
activation='relu',
name='GCN_%s' % i)
#expand_dim = Lambda(lambda x: K.expand_dims(x, axis = 1))
# sentence_output1 = conv(sentence_output)
sentence_output2 = gcn([sentence_output] + A_in)
H = Dropout(args.dropout_prob)(sentence_output2)
Global_graph = MyMeanPool(axis=1, smask=mask)(H)
node_num = H.shape.as_list()[1]
Gg = expand_dim(Global_graph)
GG = Lambda(lambda x: K.tile(x, [1, node_num, 1]))(Gg)
# code.interact(local=locals())
if args.use_cnn:
conv = Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='CNN_%s'%i)
sentence_output1 = conv(sentence_output)
# sentence_output = Concatenate()([sentence_output_1, GG, H])
if args.use_meanpool:
sentence_output = Concatenate()([sentence_output_1, GG, H])
else:
sentence_output = Concatenate()([sentence_output_1, H])
else:
if args.use_meanpool:
sentence_output = Concatenate()([GG, H])
else:
sentence_output = H
if args.use_bert_cls == 2 and args.use_bert:
sentence_output = Concatenate()([sentence_output, bert_cls])
sentence_output = share_gcn_dense(sentence_output)
if args.use_doc:
doc_output_1 = conv(doc_output_1)
doc_output_2 = conv(doc_output_2)
word_embeddings = Concatenate()([word_embeddings, sentence_output])
init_shared_features = sentence_output
#######################################
# Define task-specific layers
#######################################
# AE specific layers
aspect_cnn = Sequential()
for a in xrange(args.aspect_layers):
print 'Aspect extraction layer %s' % a
aspect_cnn.add(Dropout(args.dropout_prob))
aspect_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='aspect_cnn_%s'%a))
aspect_dense = Dense(nb_class, activation='softmax', name='aspect_dense')
aspect_dense_ = Dense(nb_class, name='aspect_dense_')
# AS specific layers
sentiment_cnn = Sequential()
for b in xrange(args.senti_layers):
print 'Sentiment classification layer %s' % b
sentiment_cnn.add(Dropout(args.dropout_prob))
sentiment_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='sentiment_cnn_%s'%b))
sentiment_att = Self_attention(args.use_opinion, name='sentiment_att')
sentiment_dense = Dense(3, activation='softmax', name='sentiment_dense')
# sentiment_dense_ = Dense(3, name='sentiment_dense_')
if args.use_doc:
# DS specific layers
doc_senti_cnn = Sequential()
for c in xrange(args.doc_senti_layers):
print 'Document-level sentiment layers %s' % c
doc_senti_cnn.add(Dropout(args.dropout_prob))
doc_senti_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='doc_sentiment_cnn_%s'%c))
doc_senti_att = Attention(name='doc_senti_att')
doc_senti_dense = Dense(3, name='doc_senti_dense')
# The reason not to use the default softmax is that it reports errors when input_dims=2 due to
# compatibility issues between the tf and keras versions used.
softmax = Lambda(lambda x: K.tf.nn.softmax(x),
name='doc_senti_softmax')
# DD specific layers
doc_domain_cnn = Sequential()
for d in xrange(args.doc_domain_layers):
print 'Document-level domain layers %s' % d
doc_domain_cnn.add(Dropout(args.dropout_prob))
doc_domain_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='doc_domain_cnn_%s'%d))
doc_domain_att = Attention(name='doc_domain_att')
doc_domain_dense = Dense(1,
activation='sigmoid',
name='doc_domain_dense')
# re-encoding layer
enc = Dense(300, activation='relu', name='enc')
####################################################
# aspect-level operations involving message passing
####################################################
for i in xrange(args.interactions + 1):
print 'Interaction number ', i
aspect_output = sentence_output
sentiment_output = sentence_output
# note that the aspet-level data will also go through the doc-level models
doc_senti_output = sentence_output
doc_domain_output = sentence_output
### AE ###
if args.aspect_layers > 0:
aspect_output = aspect_cnn(aspect_output)
# concate word embeddings and task-specific output for prediction
aspect_output = Concatenate()([word_embeddings, aspect_output])
if args.use_bert_cls == 3 and args.use_bert:
aspect_output = Concatenate()([aspect_output, bert_cls])
aspect_output = Dropout(args.dropout_prob)(aspect_output)
aspect_probs = aspect_dense(aspect_output)
use_crf = 0
if use_crf:
aspect_prob = aspect_dense_(aspect_output)
aspect_crf = CRF(nb_class, sparse_target=True) # False
aspect_crf_output = aspect_crf(aspect_prob)
### AS ###
if args.senti_layers > 0:
sentiment_output = sentiment_cnn(sentiment_output)
sentiment_output = sentiment_att(
[sentiment_output, op_label_input, aspect_probs, p_gold_op])
sentiment_output = Concatenate()(
[init_shared_features, sentiment_output])
if args.use_bert_cls == 3 and args.use_bert:
sentiment_output = Concatenate()([sentiment_output, bert_cls])
sentiment_output = Dropout(args.dropout_prob)(sentiment_output)
sentiment_probs = sentiment_dense(sentiment_output)
# use_crf = 0
# if use_crf:
# sentiment_prob = sentiment_dense_(sentiment_output)
# senti_crf = CRF(3, sparse_target=False) # False
# senti_crf_output = senti_crf(sentiment_prob)
if args.use_doc:
### DS ###
if args.doc_senti_layers > 0:
doc_senti_output = doc_senti_cnn(doc_senti_output)
# output attention weights with two activation functions
senti_att_weights_softmax, senti_att_weights_sigmoid = doc_senti_att(
doc_senti_output)
# reshape the sigmoid attention weights, will be used in message passing
senti_weights = Lambda(lambda x: K.expand_dims(x, axis=-1))(
senti_att_weights_sigmoid)
doc_senti_output = WeightedSum()(
[doc_senti_output, senti_att_weights_softmax])
doc_senti_output = Dropout(args.dropout_prob)(doc_senti_output)
doc_senti_output = doc_senti_dense(doc_senti_output)
doc_senti_probs = softmax(doc_senti_output)
# reshape the doc-level sentiment predictions, will be used in message passing
doc_senti_probs = Lambda(lambda x: K.expand_dims(x, axis=-2))(
doc_senti_probs)
doc_senti_probs = Lambda(lambda x: K.repeat_elements(
x, overall_maxlen, axis=1))(doc_senti_probs)
### DD ###
if args.doc_domain_layers > 0:
doc_domain_output = doc_domain_cnn(doc_domain_output)
domain_att_weights_softmax, domain_att_weights_sigmoid = doc_domain_att(
doc_domain_output)
domain_weights = Lambda(lambda x: K.expand_dims(x, axis=-1))(
domain_att_weights_sigmoid)
doc_domain_output = WeightedSum()(
[doc_domain_output, domain_att_weights_softmax])
doc_domain_output = Dropout(args.dropout_prob)(doc_domain_output)
doc_domain_probs = doc_domain_dense(doc_domain_output)
# update sentence_output for the next iteration
sentence_output = Concatenate()([
sentence_output, aspect_probs, sentiment_probs,
doc_senti_probs, senti_weights, domain_weights
])
else:
# update sentence_output for the next iteration
#sentence_output = Concatenate()([sentence_output, aspect_probs, sentiment_probs])
if args.use_prob:
sentence_output = Concatenate()(
[sentence_output, aspect_probs, sentiment_probs])
else:
sentence_output = Concatenate()(
[sentence_output, aspect_output, sentiment_output])
if args.use_bert_cls == 4 and args.use_bert:
sentence_output = Concatenate()([sentence_output, bert_cls])
sentence_output = enc(sentence_output)
use_crf = 0
if use_crf:
aspect_prob = aspect_dense_(aspect_output)
aspect_crf = CRF(nb_class, sparse_target=True) # False
aspect_crf_output = aspect_crf(aspect_prob)
aspect_model = Model(inputs=[sentence_input] + A_in +
[op_label_input] + [p_gold_op],
outputs=[aspect_crf_output, sentiment_probs])
else:
if args.use_bert:
aspect_model = Model(inputs=[sentence_input] + A_in +
[op_label_input] + [p_gold_op] + [bert_input],
outputs=[aspect_probs, sentiment_probs])
else:
aspect_model = Model(inputs=[sentence_input] + A_in +
[op_label_input] + [p_gold_op],
outputs=[aspect_probs, sentiment_probs])
####################################################
# doc-level operations without message passing
####################################################
if args.use_doc:
if args.doc_senti_layers > 0:
doc_output_1 = doc_senti_cnn(doc_output_1)
att_1, _ = doc_senti_att(doc_output_1)
doc_output_1 = WeightedSum()([doc_output_1, att_1])
doc_output_1 = Dropout(args.dropout_prob)(doc_output_1)
doc_output_1 = doc_senti_dense(doc_output_1)
doc_prob_1 = softmax(doc_output_1)
if args.doc_domain_layers > 0:
doc_output_2 = doc_domain_cnn(doc_output_2)
att_2, _ = doc_domain_att(doc_output_2)
doc_output_2 = WeightedSum()([doc_output_2, att_2])
doc_output_2 = Dropout(args.dropout_prob)(doc_output_2)
doc_prob_2 = doc_domain_dense(doc_output_2)
doc_model = Model(inputs=[doc_input_1, doc_input_2],
outputs=[doc_prob_1, doc_prob_2])
else:
doc_model = None
####################################################
# initialize word embeddings
####################################################
logger.info('Initializing lookup table')
# Load pre-trained word vectors.
# To save the loading time, here we load from the extracted subsets of the original embeddings,
# which only contains the embeddings of words in the vocab.
if args.use_doc:
emb_path_gen = '../glove/%s_.txt' % (args.domain)
emb_path_domain = '../domain_specific_emb/%s_.txt' % (args.domain)
else:
emb_path_gen = '../glove/%s.txt' % (args.domain)
emb_path_domain = '../domain_specific_emb/%s.txt' % (args.domain)
# Load pre-trained word vectors from the orginal large files
# If you are loading from ssd, the process would only take 1-2 mins
# If you are loading from hhd, the process would take a few hours at first try,
# and would take 1-2 mins in subsequent repeating runs (due to cache performance).
# emb_path_gen = '../glove.840B.300d.txt'
# if args.domain == 'lt':
# emb_path_domain = '../laptop_emb.vec'
# else:
# emb_path_domain = '../restaurant_emb.vec'
aspect_model.get_layer('word_emb').set_weights(
init_emb(
aspect_model.get_layer('word_emb').get_weights(), vocab,
emb_path_gen, emb_path_domain))
logger.info(' Done')
## Optimizaer algorithm
#
from optimizers import get_optimizer
optimizer = get_optimizer(args)
if args.use_doc == 1 and args.interactions > 0:
# fix the document-specific parameters when updating aspect model
aspect_model.get_layer('doc_senti_att').trainable = False
aspect_model.get_layer('doc_senti_dense').trainable = False
aspect_model.get_layer('doc_domain_att').trainable = False
if use_crf:
aspect_model.compile(
optimizer=optimizer,
loss=[aspect_crf.loss_function, 'categorical_crossentropy'],
loss_weights=[1., 1.])
else:
aspect_model.compile(
optimizer=optimizer,
loss=['categorical_crossentropy', 'categorical_crossentropy'],
loss_weights=[1., 1.])
#aspect_model.summary()
if args.use_doc == 1:
doc_model.get_layer('doc_senti_att').trainable = True
doc_model.get_layer('doc_senti_dense').trainable = True
doc_model.get_layer('doc_domain_att').trainable = True
doc_model.get_layer('doc_domain_dense').trainable = True
if args.use_doc:
doc_model.compile(
optimizer=optimizer,
loss=['categorical_crossentropy', 'binary_crossentropy'],
loss_weights=[1., 1.],
metrics=['categorical_accuracy', 'accuracy'])
return aspect_model, doc_model
def create_model(args, vocab, nb_class, overall_maxlen, doc_maxlen_1,
doc_maxlen_2):
# Funtion that initializes word embeddings
def init_emb(emb_matrix, vocab, emb_file_gen, emb_file_domain):
print 'Loading pretrained general word embeddings and domain word embeddings ...'
counter_gen = 0.
pretrained_emb = open(emb_file_gen)
for line in pretrained_emb:
tokens = line.split()
if len(tokens) != 301:
continue
word = tokens[0]
vec = tokens[1:]
try:
emb_matrix[0][vocab[word]][:300] = vec
counter_gen += 1
except KeyError:
pass
if args.use_domain_emb:
counter_domain = 0.
pretrained_emb = open(emb_file_domain)
for line in pretrained_emb:
tokens = line.split()
if len(tokens) != 101:
continue
word = tokens[0]
vec = tokens[1:]
try:
emb_matrix[0][vocab[word]][300:] = vec
counter_domain += 1
except KeyError:
pass
pretrained_emb.close()
logger.info(
'%i/%i word vectors initialized by general embeddings (hit rate: %.2f%%)'
% (counter_gen, len(vocab), 100 * counter_gen / len(vocab)))
if args.use_domain_emb:
logger.info(
'%i/%i word vectors initialized by domain embeddings (hit rate: %.2f%%)'
% (counter_domain, len(vocab),
100 * counter_domain / len(vocab)))
return emb_matrix
# Build model
logger.info('Building model ...')
print 'Building model ...'
print '\n\n'
vocab_size = len(vocab)
###################################
# Inputs
###################################
print 'Input layer'
# sequence of token indices for aspect-level data
sentence_input = Input(shape=(overall_maxlen, ),
dtype='int32',
name='sentence_input')
# gold opinion label for aspect-level data.
op_label_input = Input(shape=(overall_maxlen, 3),
dtype=K.floatx(),
name='op_label_input')
# probability of sending gold opinion labels at opinion transmission step
p_gold_op = Input(shape=(overall_maxlen, ),
dtype=K.floatx(),
name='p_gold_op')
A_in = Input(shape=(overall_maxlen, overall_maxlen),
dtype=K.floatx(),
name='A_input')
if args.use_doc:
# doc_input_1 denotes the data for sentiment classification
# doc_input_2 denotes the data for domain classification
doc_input_1 = Input(shape=(doc_maxlen_1, ),
dtype='int32',
name='doc_input_1')
doc_input_2 = Input(shape=(doc_maxlen_2, ),
dtype='int32',
name='doc_input_2')
if args.use_bert:
bert_input = Input(shape=(overall_maxlen + 1, 768),
dtype=K.floatx(),
name='bert_input') # +1 denote +cls
#########################################
# Shared word embedding layer
#########################################
print 'Word embedding layer'
word_emb = Embedding(vocab_size,
args.emb_dim,
mask_zero=True,
name='word_emb')
# aspect-level inputs
word_embeddings = word_emb(sentence_input)
sentence_output = word_embeddings
# doc-level inputs
if args.use_doc:
doc_output_1 = word_emb(doc_input_1)
# we only use general embedding for domain classification
doc_output_2 = word_emb(doc_input_2)
if args.use_domain_emb:
# mask out the domain embeddings
doc_output_2 = Remove_domain_emb()(doc_output_2)
def slice(x, index):
return x[:, index, :]
def slice1(x, index):
return x[:, index:, :]
expand_dim = Lambda(lambda x: K.expand_dims(x, axis=1))
if args.use_bert:
#code.interact(local=locals())
bert_inp = Lambda(slice1, arguments={'index': 1})(bert_input)
bert_cls = Lambda(slice, arguments={'index': 0})(bert_input)
#sentence_output = Concatenate()([sentence_output, bert_inp])
# if args.use_bert_cls:
#code.interact(local=locals())
#bert_cls = bert_input[:,0,:]
node_num = sentence_output.shape.as_list()[1]
bert_cls1 = expand_dim(bert_cls)
bert_cls = Lambda(lambda x: K.tile(x, [1, node_num, 1]))(bert_cls1)
from my_layers_algo import DigiCaps, Length, Capsule
if args.use_bert_cls == 0 and args.use_bert:
sentence_output = Concatenate()([sentence_output, bert_cls])
######################################
# Shared CNN layers
######################################
for i in xrange(args.shared_layers):
print 'Shared CNN layer %s' % i
sentence_output = Dropout(args.dropout_prob)(sentence_output)
if args.use_doc:
doc_output_1 = Dropout(args.dropout_prob)(doc_output_1)
doc_output_2 = Dropout(args.dropout_prob)(doc_output_2)
if i == 0:
#conv_0 = Conv1DWithMasking(filters=args.cnn_dim/2, kernel_size=2, \
# activation='relu', padding='same', kernel_initializer=my_init, name='cnn_0_0')
conv_1 = Conv1DWithMasking(filters=args.cnn_dim/2, kernel_size=3, \
activation='relu', padding='same', kernel_initializer=my_init, name='cnn_0_1')
#conv_2 = Conv1DWithMasking(filters=args.cnn_dim/2, kernel_size=4, \
# activation='relu', padding='same', kernel_initializer=my_init, name='cnn_0_2')
conv_3 = Conv1DWithMasking(filters=args.cnn_dim/2, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='cnn_0_3')
#sentence_output_0 = conv_0(sentence_output)
sentence_output_1 = conv_1(sentence_output)
#sentence_output_2 = conv_2(sentence_output)
sentence_output_3 = conv_3(sentence_output)
#sentence_output = Concatenate()([sentence_output_0, sentence_output_1, sentence_output_2, sentence_output_3])
sentence_output = Concatenate()(
[sentence_output_1, sentence_output_3])
if args.use_doc:
#doc_output_1_0 = conv_0(doc_output_1)
doc_output_1_1 = conv_1(doc_output_1)
#doc_output_1_2 = conv_2(doc_output_1)
doc_output_1_3 = conv_3(doc_output_1)
#doc_output_1 = Concatenate()([doc_output_1_0, doc_output_1_1, doc_output_1_2, doc_output_1_3])
doc_output_1 = Concatenate()([doc_output_1_1, doc_output_1_3])
#doc_output_2_0 = conv_0(doc_output_2)
doc_output_2_1 = conv_1(doc_output_2)
#doc_output_2_2 = conv_2(doc_output_2)
doc_output_2_3 = conv_3(doc_output_2)
#doc_output_2 = Concatenate()([doc_output_2_0, doc_output_2_1, doc_output_2_2, doc_output_2_3])
doc_output_2 = Concatenate()([doc_output_2_1, doc_output_2_3])
else:
#conv = Conv1DWithMasking(filters=args.cnn_dim, kernel_size=3, \
# activation='relu', padding='same', kernel_initializer=my_init, name='cnn_3_%s'%i)
conv_ = Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='cnn_5_%s'%i)
#sentence_output1 = conv(sentence_output)
sentence_output = conv_(sentence_output)
#sentence_output = Concatenate()([sentence_output1, sentence_output2])
if args.use_doc:
doc_output_1 = conv_(doc_output_1)
doc_output_2 = conv_(doc_output_2)
word_embeddings = Concatenate()([word_embeddings, sentence_output])
init_shared_features = sentence_output
#######################################
# Define task-specific layers
#######################################
#if args.which_dual == 'dual':
# from my_layers import Conv1DWithMasking, Remove_domain_emb, Self_attention, Attention, WeightedSum, Dual_attention
#else:
# from my_layers_algo import Conv1DWithMasking, Remove_domain_emb, Self_attention, Attention, WeightedSum, Dual_attention
# AE specific layers
aspect_cnn = Sequential()
for a in xrange(args.aspect_layers):
print 'Aspect extraction layer %s' % a
aspect_cnn.add(Dropout(args.dropout_prob))
aspect_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='aspect_cnn_%s'%a))
aspect_dense = Dense(nb_class, activation='softmax', name='aspect_dense')
# OE specific layers
opinion_cnn = Sequential()
for a in xrange(args.opinion_layers):
print 'Opinion extraction layer %s' % a
opinion_cnn.add(Dropout(args.dropout_prob))
opinion_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='opinion_cnn_%s'%a))
opinion_dense = Dense(nb_class, activation='softmax', name='opinion_dense')
# AS specific layers
sentiment_cnn = Sequential()
for b in xrange(args.senti_layers):
print 'Sentiment classification layer %s' % b
sentiment_cnn.add(Dropout(args.dropout_prob))
sentiment_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='sentiment_cnn_%s'%b))
sentiment_att = Self_attention(args.use_opinion, name='sentiment_att')
sentiment_dense = Dense(3, activation='softmax', name='sentiment_dense')
aspect_dual_att = Dual_attention(name='aspect_dualatt')
opinion_dual_att = Dual_attention(name='opinion_dualatt')
sentiment_dual_att = Dual_attention(name='sentiment_dualatt')
asp_caps = Capsule(num_capsule=overall_maxlen,
A=A_in,
dim_capsule=args.capsule_dim,
routings=3,
name='asp_caps')
senti_caps = Capsule(num_capsule=overall_maxlen,
A=A_in,
dim_capsule=args.capsule_dim,
routings=3,
name='senti_caps')
opin_caps = Capsule(num_capsule=overall_maxlen,
A=A_in,
dim_capsule=args.capsule_dim,
routings=3,
name='opin_caps')
#probs = Length(name='out_caps')
if args.use_doc:
# DS specific layers
doc_senti_cnn = Sequential()
for c in xrange(args.doc_senti_layers):
print 'Document-level sentiment layers %s' % c
doc_senti_cnn.add(Dropout(args.dropout_prob))
doc_senti_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='doc_sentiment_cnn_%s'%c))
doc_senti_att = Attention(name='doc_senti_att')
doc_senti_dense = Dense(3, name='doc_senti_dense')
# The reason not to use the default softmax is that it reports errors when input_dims=2 due to
# compatibility issues between the tf and keras versions used.
softmax = Lambda(lambda x: K.tf.nn.softmax(x),
name='doc_senti_softmax')
# DD specific layers
doc_domain_cnn = Sequential()
for d in xrange(args.doc_domain_layers):
print 'Document-level domain layers %s' % d
doc_domain_cnn.add(Dropout(args.dropout_prob))
doc_domain_cnn.add(Conv1DWithMasking(filters=args.cnn_dim, kernel_size=5, \
activation='relu', padding='same', kernel_initializer=my_init, name='doc_domain_cnn_%s'%d))
doc_domain_att = Attention(name='doc_domain_att')
doc_domain_dense = Dense(1,
activation='sigmoid',
name='doc_domain_dense')
# re-encoding layer
enc = Dense(args.cnn_dim, activation='relu', name='enc')
enc_a = Dense(args.cnn_dim, activation='relu', name='enc_a')
enc_o = Dense(args.cnn_dim, activation='relu', name='enc_o')
enc_s = Dense(args.cnn_dim, activation='relu', name='enc_s')
enc_d = Dense(args.cnn_dim, activation='relu', name='enc_d')
####################################################
# aspect-level operations involving message passing
####################################################
print(sentence_output)
# sentence_output = enc(sentence_output)
aspect_output = sentence_output
opinion_output = sentence_output
sentiment_output = sentence_output
doc_senti_output = sentence_output
doc_domain_output = sentence_output
for i in xrange(args.interactions + 1):
print 'Interaction number ', i
if args.use_doc:
### DS ###
if args.doc_senti_layers > 0:
doc_senti_output = doc_senti_cnn(doc_senti_output)
# output attention weights with two activation functions
senti_att_weights_softmax, senti_att_weights_sigmoid = doc_senti_att(
doc_senti_output)
# reshape the sigmoid attention weights, will be used in message passing
senti_weights = Lambda(lambda x: K.expand_dims(x, axis=-1))(
senti_att_weights_sigmoid)
doc_senti_output1 = WeightedSum()(
[doc_senti_output, senti_att_weights_softmax])
doc_senti_output1 = Dropout(args.dropout_prob)(doc_senti_output1)
doc_senti_output1 = doc_senti_dense(doc_senti_output1)
doc_senti_probs = softmax(doc_senti_output1)
# reshape the doc-level sentiment predictions, will be used in message passing
doc_senti_probs = Lambda(lambda x: K.expand_dims(x, axis=-2))(
doc_senti_probs)
doc_senti_probs = Lambda(lambda x: K.repeat_elements(
x, overall_maxlen, axis=1))(doc_senti_probs)
### DD ###
if args.doc_domain_layers > 0:
doc_domain_output = doc_domain_cnn(doc_domain_output)
domain_att_weights_softmax, domain_att_weights_sigmoid = doc_domain_att(
doc_domain_output)
domain_weights = Lambda(lambda x: K.expand_dims(x, axis=-1))(
domain_att_weights_sigmoid)
#code.interact(local=locals())
doc_domain_output1 = WeightedSum()(
[doc_domain_output, domain_att_weights_softmax])
doc_domain_output1 = Dropout(args.dropout_prob)(doc_domain_output1)
doc_domain_probs = doc_domain_dense(doc_domain_output1)
if args.use_bert:
aspect_output = Concatenate()([aspect_output, bert_inp])
opinion_output = Concatenate()([opinion_output, bert_inp])
sentiment_output = Concatenate()([sentiment_output, bert_inp])
aspect_output = Dropout(args.dropout_prob)(aspect_output)
opinion_output = Dropout(args.dropout_prob)(opinion_output)
sentiment_output = Dropout(args.dropout_prob)(sentiment_output)
### AE ###
if args.aspect_layers > 0:
aspect_output = aspect_cnn(aspect_output)
# concate word embeddings and task-specific output for prediction
### OE ###
if args.aspect_layers > 0:
opinion_output = opinion_cnn(opinion_output)
### AS ###
if args.senti_layers > 0:
sentiment_output = sentiment_cnn(sentiment_output)
opin2asp = asp_caps([aspect_output, opinion_output])
senti2asp = asp_caps([aspect_output, sentiment_output])
asp = Concatenate()([opin2asp, senti2asp])
asp2opin = opin_caps([opinion_output, aspect_output])
senti2opin = opin_caps([opinion_output, sentiment_output])
opin = Concatenate()([asp2opin, senti2opin])
asp2senti = senti_caps([sentiment_output, aspect_output])
opin2senti = senti_caps([sentiment_output, opinion_output])
senti = Concatenate()([asp2senti, opin2senti])
#sentiment_output = sentiment_att([sentiment_output, op_label_input, opinion_probs, p_gold_op])
# aspect_output += asp
# opinion_output += opin
# sentiment_output += senti
if args.use_doc:
aspect_output = Concatenate()(
[word_embeddings, aspect_output, asp, domain_weights])
opinion_output = Concatenate()(
[word_embeddings, opinion_output, opin, domain_weights])
sentiment_output = Concatenate()([
init_shared_features, sentiment_output, senti, doc_senti_probs,
senti_weights
])
else:
aspect_output = Concatenate()(
[word_embeddings, aspect_output, asp])
opinion_output = Concatenate()(
[word_embeddings, opinion_output, opin])
sentiment_output = Concatenate()(
[init_shared_features, sentiment_output, senti])
#aspect_output = Concatenate()([init_shared_features, aspect_output])
aspect_output = Dropout(args.dropout_prob)(aspect_output)
aspect_probs = aspect_dense(aspect_output)
#opinion_output = Concatenate()([init_shared_features, opinion_output])
opinion_output = Dropout(args.dropout_prob)(opinion_output)
opinion_probs = opinion_dense(opinion_output)
#sentiment_output = Concatenate()([word_embeddings, sentiment_output])
sentiment_output = Dropout(args.dropout_prob)(sentiment_output)
sentiment_probs = sentiment_dense(sentiment_output)
# update sentence_output for the next iteration
opinion_output = Concatenate()([
opinion_output, aspect_probs, opinion_probs, sentiment_probs,
domain_weights
])
aspect_output = Concatenate()([
aspect_output, aspect_probs, opinion_probs, sentiment_probs,
domain_weights
])
sentiment_output = Concatenate()([
sentiment_output, aspect_probs, opinion_probs, sentiment_probs,
doc_senti_probs, senti_weights
])
sentence_output_ = Concatenate()([
sentence_output, aspect_probs, opinion_probs, sentiment_probs,
doc_senti_probs, senti_weights, domain_weights
])
#code.interact(local=locals())
aspect_output = enc_a(aspect_output)
opinion_output = enc_o(opinion_output)
sentiment_output = enc_s(sentiment_output)
if args.use_doc:
doc_senti_output = enc_d(sentence_output_)
doc_domain_output = enc_d(sentence_output_)
if args.use_bert:
aspect_model = Model(
inputs=[
sentence_input, A_in, op_label_input, p_gold_op, bert_input
],
outputs=[aspect_probs, opinion_probs, sentiment_probs])
else:
aspect_model = Model(
inputs=[sentence_input, A_in, op_label_input, p_gold_op],
outputs=[aspect_probs, opinion_probs, sentiment_probs])
####################################################
# doc-level operations without message passing
####################################################
if args.use_doc:
if args.doc_senti_layers > 0:
doc_output_1 = doc_senti_cnn(doc_output_1)
att_1, _ = doc_senti_att(doc_output_1)
doc_output_1 = WeightedSum()([doc_output_1, att_1])
doc_output_1 = Dropout(args.dropout_prob)(doc_output_1)
doc_output_1 = doc_senti_dense(doc_output_1)
doc_prob_1 = softmax(doc_output_1)
if args.doc_domain_layers > 0:
doc_output_2 = doc_domain_cnn(doc_output_2)
att_2, _ = doc_domain_att(doc_output_2)
doc_output_2 = WeightedSum()([doc_output_2, att_2])
doc_output_2 = Dropout(args.dropout_prob)(doc_output_2)
doc_prob_2 = doc_domain_dense(doc_output_2)
doc_model = Model(inputs=[doc_input_1, doc_input_2],
outputs=[doc_prob_1, doc_prob_2])
else:
doc_model = None
####################################################
# initialize word embeddings
####################################################
logger.info('Initializing lookup table')
# Load pre-trained word vectors.
# To save the loading time, here we load from the extracted subsets of the original embeddings,
# which only contains the embeddings of words in the vocab.
if args.use_doc:
emb_path_gen = '../glove/%s_.txt' % (args.domain)
emb_path_domain = '../domain_specific_emb/%s_.txt' % (args.domain)
else:
emb_path_gen = '../glove/%s.txt' % (args.domain)
emb_path_domain = '../domain_specific_emb/%s.txt' % (args.domain)
# Load pre-trained word vectors from the orginal large files
# If you are loading from ssd, the process would only take 1-2 mins
# If you are loading from hhd, the process would take a few hours at first try,
# and would take 1-2 mins in subsequent repeating runs (due to cache performance).
# emb_path_gen = '../glove.840B.300d.txt'
# if args.domain == 'lt':
# emb_path_domain = '../laptop_emb.vec'
# else:
# emb_path_domain = '../restaurant_emb.vec'
aspect_model.get_layer('word_emb').set_weights(
init_emb(
aspect_model.get_layer('word_emb').get_weights(), vocab,
emb_path_gen, emb_path_domain))
logger.info(' Done')
return aspect_model, doc_model
