def get_embeddings(idx2word, config):
embedding = np.random.uniform(-1 / 16, 1 / 16,
[config.vocab_size, config.embed_dim])
if config.pretrained_embed_dir:
processor = JamoProcessor()
ft = FastText.load(config.pretrained_embed_dir)
num_oov = 0
for i, vocab in enumerate(idx2word):
try:
embedding[i, :] = ft.wv[processor.word_to_jamo(vocab)]
except:
num_oov += 1
print("Pre-trained embedding loaded. Number of OOV : {} / {}".format(
num_oov, len(idx2word)))
else:
print(
"No pre-trained embedding found, initialize with random distribution"
)
return embedding
def get_embeddings(vocab_list_dir,
pretrained_embed_dir,
vocab_size,
embed_dim):
embedding = np.random.uniform(-1/16, 1/16, [vocab_size, embed_dim])
if os.path.isfile(pretrained_embed_dir) & os.path.isfile(vocab_list_dir):
with open(vocab_list_dir, "r", encoding="utf-8") as f:
vocab_list = [word.strip() for word in f if len(word)>0]
processor = JamoProcessor()
ft = FastText.load(pretrained_embed_dir)
num_oov = 0
for i, vocab in enumerate(vocab_list):
try:
embedding[i, :] = ft.wv[processor.word_to_jamo(vocab)]
except:
num_oov += 1
print("Pre-trained embedding loaded. Number of OOV : {} / {}".format(num_oov, len(vocab_list)))
else:
print("No pre-trained embedding found, initialize with random distribution")
return embedding
class FeatureExtractor:
def __init__(self):
self.tfidf_vectorizer = None
self.jamo_processor = JamoProcessor()
def tokens_diff(self, a, b):
a_tokens = a.split(" ")
b_tokens = b.split(" ")
return len(set(a_tokens) & set(b_tokens)) / max(
len(a_tokens), len(b_tokens))
def edit_distance(self, a, b):
a_jamos = self.jamo_processor.word_to_jamo(a).replace("_", "")
b_jamos = self.jamo_processor.word_to_jamo(b).replace("_", "")
return editdistance.eval(a_jamos, b_jamos)
def extract_features(self, A, B):
extracted_features = list()
for a, b in zip(A, B):
ls = [self.tokens_diff(a, b), self.edit_distance(a, b)]
extracted_features.append(ls)
return extracted_features
class FeatureExtractor:
def __init__(self):
self.infer_model, self.infer_sess = self._load_pretrained_model()
self.tfidf_char_vectorizer = pickle.load(
open("../dump/tfidf_char_vectorizer.pkl", "rb"))
self.tfidf_word_vectorizer = pickle.load(
open("../dump/tfidf_word_vectorizer_big.pkl", "rb"))
self.processor = JamoProcessor()
self.tokenizer = SentencePieceTokenizer(config)
def _load_pretrained_model(self):
base_dir = "/media/scatter/scatterdisk/reply_matching_model/runs/delstm_1024_nsrandom4_lr1e-3/"
config_dir = base_dir + "config.json"
best_model_dir = base_dir + "best_loss/best_loss.ckpt"
model_config = load_config(config_dir)
model_config.add_echo = False
preprocessor = DynamicPreprocessor(model_config)
preprocessor.build_preprocessor()
infer_config = load_config(config_dir)
setattr(infer_config, "tokenizer", "SentencePieceTokenizer")
setattr(
infer_config, "soynlp_scores",
"/media/scatter/scatterdisk/tokenizer/soynlp_scores.sol.100M.txt")
infer_preprocessor = DynamicPreprocessor(infer_config)
infer_preprocessor.build_preprocessor()
graph = tf.Graph()
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
with graph.as_default():
Model = get_model(model_config.model)
data = DataGenerator(preprocessor, model_config)
infer_model = Model(data, model_config)
infer_sess = tf.Session(config=tf_config, graph=graph)
infer_sess.run(tf.global_variables_initializer())
infer_sess.run(tf.local_variables_initializer())
infer_model.load(infer_sess, model_dir=best_model_dir)
self.infer_preprocessor = infer_preprocessor
return infer_model, infer_sess
def _batch_infer(self, batch_A, batch_B):
indexed_A, A_lengths = zip(
*[self.infer_preprocessor.preprocess(a) for a in batch_A])
indexed_B, B_lengths = zip(
*[self.infer_preprocessor.preprocess(b) for b in batch_B])
feed_dict = {
self.infer_model.input_queries: indexed_A,
self.infer_model.input_replies: indexed_B,
self.infer_model.queries_lengths: A_lengths,
self.infer_model.replies_lengths: B_lengths,
self.infer_model.dropout_keep_prob: 1,
}
A_sentence_vectors, AB_probs = self.infer_sess.run([
self.infer_model.encoding_queries, self.infer_model.positive_probs
],
feed_dict=feed_dict)
feed_dict = {
self.infer_model.input_queries: indexed_B,
self.infer_model.input_replies: indexed_A,
self.infer_model.queries_lengths: B_lengths,
self.infer_model.replies_lengths: A_lengths,
self.infer_model.dropout_keep_prob: 1,
}
B_sentence_vectors, BA_probs = self.infer_sess.run([
self.infer_model.encoding_queries, self.infer_model.positive_probs
],
feed_dict=feed_dict)
semantic_sim = [
cosine_similarity([a_vector], [b_vector])[0][0]
for a_vector, b_vector in zip(list(A_sentence_vectors),
list(B_sentence_vectors))
]
return [p[0] for p in AB_probs], [p[0] for p in BA_probs], semantic_sim
def extract_features(self, sentences_A, sentences_B):
def get_semantic_sim(A, B, batch_size=512):
length = len(A)
num_batches = (length - 1) // batch_size + 1
result = {
"ab_probs": list(),
"ba_probs": list(),
"semantic_sim": list()
}
for batch_num in range(num_batches):
start = batch_num * batch_size
end = min([(batch_num + 1) * batch_size, length])
ab_probs, ba_probs, semantic_sim = self._batch_infer(
A[start:end], B[start:end])
result["ab_probs"] += list(ab_probs)
result["ba_probs"] += list(ba_probs)
result["semantic_sim"] += semantic_sim
return result
def get_word_tfidf_sim(A, B):
word_sim = list()
for a, b in zip(A, B):
word_sim.append(
cosine_similarity(
self.tfidf_word_vectorizer.transform([a]),
self.tfidf_word_vectorizer.transform([b]))[0][0])
return {"tfidf_word_sim": word_sim}
def get_char_tfidf_sim(A, B):
char_sim = list()
for a, b in zip(A, B):
char_sim.append(
cosine_similarity(
self.tfidf_char_vectorizer.transform([a]),
self.tfidf_char_vectorizer.transform([b]))[0][0])
return {"tfidf_char_sim": char_sim}
def get_edit_distance(A, B):
edit_distance = list()
substring_ratio = list()
for a, b in zip(A, B):
a_jamos = self.processor.word_to_jamo(a).replace("_", "")
b_jamos = self.processor.word_to_jamo(b).replace("_", "")
edit_distance.append(proper_edit_distance(a_jamos, b_jamos))
substring_ratio.append(substring(a_jamos, b_jamos))
return {
"edit_distance": edit_distance,
"substring_ratio": substring_ratio
}
extracted_features = dict()
extracted_features.update(
get_semantic_sim(sentences_A, sentences_B, batch_size=512))
extracted_features.update(get_word_tfidf_sim(sentences_A, sentences_B))
extracted_features.update(get_char_tfidf_sim(sentences_A, sentences_B))
extracted_features.update(get_edit_distance(sentences_A, sentences_B))
return extracted_features