|
from huggingface_hub import from_pretrained_keras |
|
import tensorflow as tf |
|
import gradio as gr |
|
import nltk |
|
nltk.download('brown') |
|
from nltk.corpus import brown |
|
from nltk import word_tokenize |
|
nltk.download('punkt') |
|
nltk.download('stopwords') |
|
nltk.download('wordnet') |
|
from nltk.corpus import stopwords |
|
from nltk import pos_tag |
|
nltk.download('averaged_perceptron_tagger') |
|
import re |
|
import torch |
|
from transformers import BertTokenizer, BertModel, BertForMaskedLM |
|
from wordfreq import zipf_frequency |
|
import keras |
|
from keras_preprocessing.sequence import pad_sequences |
|
from tensorflow.keras.utils import to_categorical |
|
|
|
bert_model = 'bert-large-uncased' |
|
tokenizer = BertTokenizer.from_pretrained(bert_model) |
|
model = BertForMaskedLM.from_pretrained(bert_model) |
|
|
|
model_cwi = from_pretrained_keras("jaimin/CWI") |
|
|
|
stop_words_ = set(stopwords.words('english')) |
|
def cleaner(word): |
|
|
|
word = re.sub(r'((http|https)\:\/\/)?[a-zA-Z0-9\.\/\?\:@\-_=#]+\.([a-zA-Z]){2,6}([a-zA-Z0-9\.\&\/\?\:@\-_=#])*', |
|
'', word, flags=re.MULTILINE) |
|
word = re.sub('[\W]', ' ', word) |
|
word = re.sub('[^a-zA-Z]', ' ', word) |
|
return word.lower().strip() |
|
|
|
def process_input(input_text): |
|
input_text = cleaner(input_text) |
|
clean_text = [] |
|
index_list =[] |
|
input_token = [] |
|
index_list_zipf = [] |
|
for i, word in enumerate(input_text.split()): |
|
clean_text.append(word) |
|
index_list.append(i) |
|
input_token.append(word) |
|
input_padded = pad_sequences(maxlen=512, sequences=[input_token], padding="post", value=0) |
|
return input_padded, index_list, len(clean_text) |
|
|
|
def complete_missing_word(pred_binary, index_list, len_list): |
|
list_cwi_predictions = list(pred_binary[0][:len_list]) |
|
for i in index_list: |
|
list_cwi_predictions.insert(i, 0) |
|
return list_cwi_predictions |
|
|
|
|
|
def get_bert_candidates(input_text, list_cwi_predictions, numb_predictions_displayed = 10): |
|
list_candidates_bert = [] |
|
for word,pred in zip(input_text.split(), list_cwi_predictions): |
|
if (pred and (pos_tag([word])[0][1] in ['NNS', 'NN', 'VBP', 'RB', 'VBG','VBD' ])) or (zipf_frequency(word, 'en')) <3.1: |
|
replace_word_mask = input_text.replace(word, '[MASK]') |
|
text = f'[CLS]{replace_word_mask} [SEP] {input_text} [SEP] ' |
|
tokenized_text = tokenizer.tokenize(text) |
|
masked_index = [i for i, x in enumerate(tokenized_text) if x == '[MASK]'][0] |
|
indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text) |
|
segments_ids = [0]*len(tokenized_text) |
|
tokens_tensor = torch.tensor([indexed_tokens]) |
|
segments_tensors = torch.tensor([segments_ids]) |
|
|
|
with torch.no_grad(): |
|
outputs = model(tokens_tensor, token_type_ids=segments_tensors) |
|
predictions = outputs[0][0][masked_index] |
|
predicted_ids = torch.argsort(predictions, descending=True)[:numb_predictions_displayed] |
|
predicted_tokens = tokenizer.convert_ids_to_tokens(list(predicted_ids)) |
|
list_candidates_bert.append((word, predicted_tokens)) |
|
return list_candidates_bert |
|
|
|
def cwi(input_text): |
|
new_text = input_text |
|
input_padded, index_list, len_list = process_input(input_text) |
|
pred_cwi = model_cwi.predict(input_padded) |
|
pred_cwi_binary = np.argmax(pred_cwi, axis = 2) |
|
complete_cwi_predictions = complete_missing_word(pred_cwi_binary, index_list, len_list) |
|
bert_candidates = get_bert_candidates(input_text, complete_cwi_predictions) |
|
for word_to_replace, l_candidates in bert_candidates: |
|
tuples_word_zipf = [] |
|
for w in l_candidates: |
|
if w.isalpha(): |
|
tuples_word_zipf.append((w, zipf_frequency(w, 'en'))) |
|
tuples_word_zipf = sorted(tuples_word_zipf, key = lambda x: x[1], reverse=True) |
|
new_text = re.sub(word_to_replace, tuples_word_zipf[0][0], new_text) |
|
return new_text |
|
|
|
interface = gr.Interface(fn=cwi, |
|
inputs=["text"], |
|
outputs="text", |
|
title='CWI') |
|
|
|
|
|
interface.launch(inline=False) |