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)

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, None, None
    word_to_mask = random.choice(non_stop_words)
    masked_sentence = sentence.replace(word_to_mask, '[MASK]', 1)
    predictions = fill_mask(masked_sentence)
    words = [pred['score'] for pred in predictions]
    logits = [pred['token_str'] for pred in predictions]
    return masked_sentence, words, logits

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, None, None
    random.seed(10)
    word_to_mask = random.choice(non_stop_words)
    masked_sentence = sentence.replace(word_to_mask, '[MASK]', 1)
    predictions = fill_mask(masked_sentence)
    words = [pred['score'] for pred in predictions]
    logits = [pred['token_str'] for pred in predictions]
    return masked_sentence, words, logits

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, None, None
    
    max_entropy = -float('inf')
    max_entropy_word = None
    max_logits = 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
            max_logits = [pred['score'] for pred in predictions]
    
    masked_sentence = sentence.replace(max_entropy_word, '[MASK]', 1)
    words = [pred['score'] for pred in predictions]
    logits = [pred['token_str'] for pred in predictions]
    return masked_sentence, words, logits

# 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)

non_melting_points = [(1, 'Jewish'), (2, 'messages'), (3, 'stab')]
a, b, c = high_entropy_words("A former Cornell University student was sentenced to 21 months in prison on Monday after admitting that he had posted a series of online messages last fall in which he threatened to stab, rape and behead Jewish people", non_melting_points)
print(f"logits type: {type(b)}")
print(f"logits content: {b}")