def build_datasets(reduced=False, columns=None):
random.seed(0)
vectors = WordVectors(load=True)
build_dataset(vectors, "train", reduced=reduced, columns=columns, tune_fraction=0.15)
build_dataset(vectors, "train", reduced=reduced, columns=columns)
build_dataset(vectors, "dev", reduced=reduced, columns=columns)
build_dataset(vectors, "test", reduced=reduced, columns=columns)
from word_vectors import WordVectors
import numpy as np
# syn_final.npy: storing word embeddings, numpy array of shape [vocab_size, hidden_size]
# 'vocab.txt': text file storing words in vocabulary, one word per line
query = ','
num_similar_words = 10
syn0_final = np.load('syn0_final.npy')
vocab_words = []
with open('vocab.txt') as f:
vocab_words = [l.strip() for l in f]
wv = WordVectors(syn0_final, vocab_words)
print(wv.most_similar(query, num_similar_words))
from word_vectors import WordVectors
import numpy as np
import train
import evaluate
import sys
# dataset = Dataset("SST")
dataset = Dataset("SST_phrase")
# dataset = Dataset("IMDB", preprocess=True)
# dataset = Dataset("MR", preprocess=True)
# load pretrained word vectors
pretrained_vectors = []
for index, type_ in enumerate(config['word_vector_type']):
pretrained_vectors.append(
WordVectors(type_, config["pretrained_vectors"][index]))
print("loaded vectors {}".format(config['word_vector_type'][index]))
# data = dataset.cv_split(index=5)
data = dataset.cv_split(index=2)
# # data = dataset.cv_split(index=5)
# SST splits (take only train and test)
# 1 = train
# 2 = test
# 3 = dev
t1 = dataset.cv_split(index=2)
# t1 = dataset.cv_split(index=3)
t2 = dataset.cv_split(index=1)
data = [t2[2], t2[3], t1[2], t1[3]]
def main():
"""
The user can provide the location of the config file as an argument.
If no location is specified, the default config file
(experiment_parameters.cfg) is used.
"""
# load pretrained word vectors
language = "english"
pretrained_vectors = []
for index, type_ in enumerate(config['word_vector_type']):
pretrained_vectors.append(
WordVectors(type_, config["pretrained_vectors"][index]))
print("loaded vectors {}".format(config['word_vector_type'][index]))
# load from evaulation path
type_ = "from_model"
# words_path = "./runs/1496739336/evaluations/1496933648/words_embds.csv"
# words_path = "./runs/1496969351/best_snaps/../evaluations/1497016586/words_embds.csv"
# words_path = "./runs/1497028147/best_snaps/../evaluations/1497092785/words_embds.csv"
# words_path = "./runs/1497313304/best_snaps/../evaluations/1497438814/words_embds.csv"
# words_path = "./runs/1497517616/best_snaps/../evaluations/1497862652/words_embds.csv"
# words_path = "./runs/1497815260/best_snaps/../evaluations/1498091033/words_embds.csv"
# words_path = "./runs/1497894091/best_snaps/../evaluations/1498148645/words_embds.csv"
words_path = "./runs/1498124433/best_snaps/../evaluations/1498157705/words_embds.csv"
pretrained_vectors.append(WordVectors(type_, words_path))
config['word_vector_type'].append(type_)
flag_ = True
for vec_num_, word_vectors in enumerate(pretrained_vectors):
# print (word_vectors.dictionary)
# print (word_vectors.vectors[0])
# print (word_vectors.word_to_index["sermonize"], "seromnize")
print("\n============= Evaluating word vectors: {} for language: {}"
" =============\n".format(config['word_vector_type'][vec_num_],
(language)))
simlex_score, simlex_coverage = evaluate_similarity(
word_vectors, language)
print("SimLex-999 score and coverage:", simlex_score, simlex_coverage)
# sys.exit()
# WordSim Validation scores:
c1, cov1 = evaluate_similarity(word_vectors, language, source=language)
c2, cov2 = evaluate_similarity(word_vectors,
language,
source=language + "-sim")
c3, cov3 = evaluate_similarity(word_vectors,
language,
source=language + "-rel")
print("WordSim overall score and coverage:", c1, cov1)
print("WordSim Similarity score and coverage:", c2, cov2)
print("WordSim Relatedness score and coverage:", c3, cov3, "\n")
men_score, men_coverage = evaluate_similarity(word_vectors,
language,
source="MEN")
print("MEN score and coverage:", men_score, men_coverage, "\n")
# SimVerb
sim_relations = [
"SYNONYMS", "ANTONYMS", "HYPER/HYPONYMS", "COHYPONYMS", "NONE",
"ALL"
]
for relation_ in sim_relations:
sv_score, sv_coverage = evaluate_similarity(word_vectors,
relation_,
source="SimVerb")
print("SimVerb {} score and coverage: {} {}".format(
relation_, sv_score, sv_coverage))
print("\n")
rw_score, rw_coverage = evaluate_similarity(word_vectors,
language,
source="rw")
print("RW score and coverage:", rw_score, rw_coverage)
if flag_:
print(
"\n======== Evaluating only words from specified dictionary ========\n"
)
for vec_num_, word_vectors in enumerate(pretrained_vectors[:-1]):
dic_ = pretrained_vectors[-1].word_to_index
simlex_score, simlex_coverage = evaluate_similarity(
word_vectors, language, dictionary_=dic_)
print("SimLex-999 score and coverage:", simlex_score,
simlex_coverage)
# sys.exit()
# WordSim Validation scores:
c1, cov1 = evaluate_similarity(word_vectors,
language,
source=language,
dictionary_=dic_)
c2, cov2 = evaluate_similarity(word_vectors,
language,
source=language + "-sim",
dictionary_=dic_)
c3, cov3 = evaluate_similarity(word_vectors,
language,
source=language + "-rel",
dictionary_=dic_)
print("WordSim overall score and coverage:", c1, cov1)
print("WordSim Similarity score and coverage:", c2, cov2)
print("WordSim Relatedness score and coverage:", c3, cov3, "\n")
men_score, men_coverage = evaluate_similarity(word_vectors,
language,
source="MEN",
dictionary_=dic_)
print("MEN score and coverage:", men_score, men_coverage, "\n")
# SimVerb
sim_relations = [
"SYNONYMS", "ANTONYMS", "HYPER/HYPONYMS", "COHYPONYMS", "NONE",
"ALL"
]
for relation_ in sim_relations:
sv_score, sv_coverage = evaluate_similarity(word_vectors,
relation_,
source="SimVerb",
dictionary_=dic_)
print("SimVerb {} score and coverage: {} {}".format(
relation_, sv_score, sv_coverage))
print("\n")
rw_score, rw_coverage = evaluate_similarity(word_vectors,
language,
source="rw",
dictionary_=dic_)
print("RW score and coverage:", rw_score, rw_coverage)