lib/comparison.py

from transformers import BertTokenizer, BertModel
import torch
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np

# Load BERT tokenizer and model
bert_model_name = "bert-base-uncased"
tokenizer = BertTokenizer.from_pretrained(bert_model_name)
model = BertModel.from_pretrained(bert_model_name)
model.eval()  # Set to evaluation mode

# Function to obtain BERT embeddings
def get_bert_embeddings(texts):
    """Obtain BERT embeddings for a list of texts."""
    embeddings = []
    with torch.no_grad():
        for text in texts:
            inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True)
            outputs = model(**inputs)
            # Take the mean of token embeddings as the sentence embedding
            embedding = outputs.last_hidden_state.mean(dim=1).squeeze().numpy()
            embeddings.append(embedding)
    return np.array(embeddings)

# Compute similarity matrices over embeddings
def compute_similarity(embeddings1, embeddings2):
    """Compute pairwise cosine similarity between two sets of embeddings."""
    return cosine_similarity(embeddings1, embeddings2)

# Compare a paragraph with a list of other paragraphs
def compare_summaries(paragraph, paragraphs):
    """
    Compare a single paragraph with a list of summaries, 
    and return the most similar summary along with the similarity score.
    """
    # Get embeddings for the paragraph and the list of summaries
    paragraph_embedding = get_bert_embeddings([paragraph])[0]  # Single paragraph embedding
    summaries_embeddings = get_bert_embeddings(paragraphs)      # Embeddings for list of paragraphs

    # Compute similarity between the paragraph and each summary
    similarities = compute_similarity([paragraph_embedding], summaries_embeddings)[0]

    # Find the most similar summary
    most_similar_index = np.argmax(similarities)               # Get index of most similar summary
    most_similar_summary = paragraphs[most_similar_index]       # Corresponding summary
    similarity_score = similarities[most_similar_index]        # Similarity score

    return most_similar_summary