masking_methods.py

from transformers import AutoTokenizer, AutoModelForMaskedLM
from transformers import pipeline
import random
from nltk.corpus import stopwords
import math

# Masking Model
def mask_non_stopword(sentence):
    stop_words = set(stopwords.words('english'))
    words = sentence.split()
    non_stop_words = [word for word in words if word.lower() not in stop_words]
    if not non_stop_words:
        return sentence
    word_to_mask = random.choice(non_stop_words)
    masked_sentence = sentence.replace(word_to_mask, '[MASK]', 1)
    return masked_sentence

def mask_non_stopword_pseudorandom(sentence):
    stop_words = set(stopwords.words('english'))
    words = sentence.split()
    non_stop_words = [word for word in words if word.lower() not in stop_words]
    if not non_stop_words:
        return sentence
    random.seed(10)
    word_to_mask = random.choice(non_stop_words)
    masked_sentence = sentence.replace(word_to_mask, '[MASK]', 1)
    return masked_sentence

def high_entropy_words(sentence, non_melting_points):
    stop_words = set(stopwords.words('english'))
    words = sentence.split()
    
    non_melting_words = set()
    for _, point in non_melting_points:
        non_melting_words.update(point.lower().split())
    
    candidate_words = [word for word in words if word.lower() not in stop_words and word.lower() not in non_melting_words]
    
    if not candidate_words:
        return sentence
    
    max_entropy = -float('inf')
    max_entropy_word = None
    
    for word in candidate_words:
        masked_sentence = sentence.replace(word, '[MASK]', 1)
        predictions = fill_mask(masked_sentence)
        
        # Calculate entropy based on top 5 predictions
        entropy = -sum(pred['score'] * math.log(pred['score']) for pred in predictions[:5])
        
        if entropy > max_entropy:
            max_entropy = entropy
            max_entropy_word = word
    
    return sentence.replace(max_entropy_word, '[MASK]', 1)


# Load tokenizer and model for masked language model
tokenizer = AutoTokenizer.from_pretrained("bert-large-cased-whole-word-masking")
model = AutoModelForMaskedLM.from_pretrained("bert-large-cased-whole-word-masking")
fill_mask = pipeline("fill-mask", model=model, tokenizer=tokenizer)

def mask(sentence):
    predictions = fill_mask(sentence)
    masked_sentences = [predictions[i]['sequence'] for i in range(len(predictions))]
    return masked_sentences