def preprocess_BNLI_data(input_file, targetdir, worddict, labeldict):
"""
Preprocess the BNLI data set so it can be used to test a model trained
on SNLI.
Args:
inputdir: The path to the file containing the Breaking NLI (BNLI) data.
target_dir: The path to the directory where the preprocessed Breaking
NLI data must be saved.
worddict: The path to the pickled worddict used for preprocessing the
training data on which models were trained before being tested on
BNLI.
labeldict: The dict of labels used for the training data on which
models were trained before being tested on BNLI.
"""
if not os.path.exists(targetdir):
os.makedirs(targetdir)
output_file = os.path.join(targetdir, "bnli.txt")
print(20 * "=", " Preprocessing Breaking NLI data set ", 20 * "=")
print("\t* Tranforming jsonl data to txt...")
jsonl_to_txt(input_file, output_file)
preprocessor = Preprocessor(labeldict=labeldict)
with open(worddict, 'rb') as pkl:
wdict = pickle.load(pkl)
preprocessor.worddict = wdict
print("\t* Reading txt data...")
data = preprocessor.read_data(output_file)
print("\t* Transforming words in premises and hypotheses to indices...")
transformed_data = preprocessor.transform_to_indices(data)
print("\t* Saving result...")
with open(os.path.join(targetdir, "bnli_data.pkl"), 'wb') as pkl_file:
pickle.dump(transformed_data, pkl_file)
def preprocess_quora_data(inputdir,
embeddings_file,
targetdir,
lowercase=False,
ignore_punctuation=False,
num_words=None,
stopwords=[],
labeldict={},
bos=None,
eos=None):
"""
Preprocess the data from the SNLI corpus so it can be used by the
ESIM model.
Compute a worddict from the train set, and transform the words in
the sentences of the corpus to their indices, as well as the labels.
Build an embedding matrix from pretrained word vectors.
The preprocessed data is saved in pickled form in some target directory.
Args:
inputdir: The path to the directory containing the NLI corpus.
embeddings_file: The path to the file containing the pretrained
word vectors that must be used to build the embedding matrix.
targetdir: The path to the directory where the preprocessed data
must be saved.
lowercase: Boolean value indicating whether to lowercase the premises
and hypotheseses in the input data. Defautls to False.
ignore_punctuation: Boolean value indicating whether to remove
punctuation from the input data. Defaults to False.
num_words: Integer value indicating the size of the vocabulary to use
for the word embeddings. If set to None, all words are kept.
Defaults to None.
stopwords: A list of words that must be ignored when preprocessing
the data. Defaults to an empty list.
bos: A string indicating the symbol to use for beginning of sentence
tokens. If set to None, bos tokens aren't used. Defaults to None.
eos: A string indicating the symbol to use for end of sentence tokens.
If set to None, eos tokens aren't used. Defaults to None.
"""
if not os.path.exists(targetdir):
os.makedirs(targetdir)
# Retrieve the train, dev and test data files from the dataset directory.
train_file = ""
dev_file = ""
test_file = ""
for file in os.listdir(inputdir):
if fnmatch.fnmatch(file, "train.tsv"):
train_file = file
elif fnmatch.fnmatch(file, "dev.tsv"):
dev_file = file
elif fnmatch.fnmatch(file, "test.tsv"):
test_file = file
# -------------------- Train data preprocessing -------------------- #
preprocessor = Preprocessor(lowercase=lowercase,
ignore_punctuation=ignore_punctuation,
num_words=num_words,
stopwords=stopwords,
labeldict=labeldict,
bos=bos,
eos=eos)
print(20*"=", " Preprocessing train set ", 20*"=")
print("\t* Reading data...")
data = preprocessor.read_data_quora(os.path.join(inputdir, train_file)) #read_data_quora
with open(os.path.join(targetdir, "elmo_train_data.pkl"), "wb") as pkl_file:
pickle.dump(data, pkl_file)
print("\t* Computing worddict and saving it...")
preprocessor.build_worddict(data)
with open(os.path.join(targetdir, "worddict.pkl"), "wb") as pkl_file:
pickle.dump(preprocessor.worddict, pkl_file)
print("\t* Transforming words in premises and hypotheses to indices...")
transformed_data = preprocessor.transform_to_indices(data)
print("\t* Saving result...")
with open(os.path.join(targetdir, "train_data.pkl"), "wb") as pkl_file:
pickle.dump(transformed_data, pkl_file)
# -------------------- Validation data preprocessing -------------------- #
print(20*"=", " Preprocessing dev set ", 20*"=")
print("\t* Reading data...")
data = preprocessor.read_data_quora(os.path.join(inputdir, dev_file))
with open(os.path.join(targetdir, "elmo_dev_data.pkl"), "wb") as pkl_file:
pickle.dump(data, pkl_file)
print("\t* Transforming words in premises and hypotheses to indices...")
transformed_data = preprocessor.transform_to_indices(data)
print("\t* Saving result...")
with open(os.path.join(targetdir, "dev_data.pkl"), "wb") as pkl_file:
pickle.dump(transformed_data, pkl_file)
# -------------------- Test data preprocessing -------------------- #
print(20*"=", " Preprocessing test set ", 20*"=")
print("\t* Reading data...")
data = preprocessor.read_data_quora(os.path.join(inputdir, test_file))
with open(os.path.join(targetdir, "elmo_test_data.pkl"), "wb") as pkl_file:
pickle.dump(data, pkl_file)
print("\t* Transforming words in premises and hypotheses to indices...")
transformed_data = preprocessor.transform_to_indices(data)
print("\t* Saving result...")
with open(os.path.join(targetdir, "test_data.pkl"), "wb") as pkl_file:
pickle.dump(transformed_data, pkl_file)
# -------------------- Embeddings preprocessing -------------------- #
print(20*"=", " Preprocessing embeddings ", 20*"=")
print("\t* Building embedding matrix and saving it...")
embed_matrix = preprocessor.build_embedding_matrix(embeddings_file)
with open(os.path.join(targetdir, "embeddings.pkl"), "wb") as pkl_file:
pickle.dump(embed_matrix, pkl_file)