AnnaWegmann/Highlight-Paraphrases-in-Dialog

Model was created as described in https://arxiv.org/abs/2404.06670 , this is the best DeBERTa AGGREGATED model. See also the GitHub repository.

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

class ParaphraseHighlighter:
    def __init__(self, model_name="AnnaWegmann/Highlight-Paraphrases-in-Dialog"):
        # Load the tokenizer and model
        self.tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
        self.model = AutoModelForTokenClassification.from_pretrained(model_name)
        
        # Get the label id for 'LABEL_1'
        self.label2id = self.model.config.label2id
        self.label_id = self.label2id['LABEL_1']
    
    def highlight_paraphrase(self, text1, text2):
        # Tokenize the inputs with the tokenizer
        encoding = self.tokenizer(text1, text2, return_tensors="pt", padding=True, truncation=True)
        
        outputs = self.model(**encoding)
        logits = outputs.logits  # Shape: (batch_size, sequence_length, num_labels)
        # Apply softmax to get probabilities, automatically places [SEP] token
        probs = torch.nn.functional.softmax(logits, dim=-1)  # Shape: (batch_size, sequence_length, num_labels)
        
        # Convert token IDs back to tokens
        tokens = self.tokenizer.convert_ids_to_tokens(encoding["input_ids"][0])        
        # Get word IDs to map tokens to words
        word_ids = encoding.word_ids(batch_index=0)
        # Get sequence IDs to know which text the token belongs to
        sequence_ids = encoding.sequence_ids(batch_index=0)
        
        # Collect words and probabilities for each text
        words_text1 = []
        words_text2 = []
        probs_text1 = []
        probs_text2 = []
        
        previous_word_idx = None
        
        # For determining if there are high-probability words in both texts
        has_high_prob_text1 = False
        has_high_prob_text2 = False
        
        for idx, (word_idx, seq_id) in enumerate(zip(word_ids, sequence_ids)):
            if word_idx is None:
                # Skip special tokens like [CLS], [SEP], [PAD]
                continue

            if word_idx != previous_word_idx:
                # Start of a new word
                word_tokens = [tokens[idx]]

                # Get the probability for LABEL_1 for the first token of the word
                prob_LABEL_1 = probs[0][idx][self.label_id].item()

                # Collect subsequent tokens belonging to the same word
                j = idx + 1
                while j < len(word_ids) and word_ids[j] == word_idx:
                    word_tokens.append(tokens[j])
                    j += 1

                # Reconstruct the word
                word = self.tokenizer.convert_tokens_to_string(word_tokens).strip()

                # Check if probability >= 0.5 to uppercase
                if prob_LABEL_1 >= 0.5:
                    word_display = word.upper()
                    if seq_id == 0:
                        has_high_prob_text1 = True
                    elif seq_id == 1:
                        has_high_prob_text2 = True
                else:
                    word_display = word

                # Append the word and probability to the appropriate list
                if seq_id == 0:
                    words_text1.append(word_display)
                    probs_text1.append(prob_LABEL_1)
                elif seq_id == 1:
                    words_text2.append(word_display)
                    probs_text2.append(prob_LABEL_1)
                else:
                    # Should not happen
                    pass

            previous_word_idx = word_idx
        
        # Determine if there are words in both texts with prob >= 0.5
        if has_high_prob_text1 and has_high_prob_text2:
            print("is a paraphrase")
        else:
            print("is not a paraphrase")
        
        # Function to format and align words and probabilities
        def print_aligned(words, probs):
            # Determine the maximum length of words for formatting
            max_word_length = max(len(word) for word in words)
            # Create format string for alignment
            format_str = f'{{:<{max_word_length}}}'
            # Print words
            for word in words:
                print(format_str.format(word), end=' ')
            print()
            # Print probabilities aligned below words
            for prob in probs:
                prob_str = f"{prob:.2f}"
                print(format_str.format(prob_str), end=' ')
            print('\n')
        
        # Print text1's words and probabilities aligned
        print("\nSpeaker 1:")
        print_aligned(words_text1, probs_text1)
        
        # Print text2's words and probabilities aligned
        print("Speaker 2:")
        print_aligned(words_text2, probs_text2)
        
# Example usage
highlighter = ParaphraseHighlighter()
text1 = "And it will be my 20th time in doing it as a television commentator from Rome so."
text2 = "Yes, you've been doing this for a while now."
highlighter.highlight_paraphrase(text1, text2)

should return

is a paraphrase

Speaker 1:
And         IT          will        BE          MY          20TH        TIME        IN          DOING       IT          as          a           TELEVISION  COMMENTATOR from        Rome        so.         
0.15        0.54        0.49        0.56        0.74        0.83        0.77        0.75        0.78        0.76        0.44        0.45        0.52        0.52        0.30        0.37        0.21        

Speaker 2:
Yes,   YOU'VE BEEN   DOING  THIS   FOR    A      WHILE  NOW.   
0.12   0.79   0.78   0.82   0.82   0.69   0.70   0.72   0.66   

For comments or questions reach out to Anna (a.m.wegmann @ uu.nl) or raise an issue on GitHub.

If you find this model helpful, consider citing our paper:

@inproceedings{wegmann-etal-2024-whats,
    title = "What{'}s Mine becomes Yours: Defining, Annotating and Detecting Context-Dependent Paraphrases in News Interview Dialogs",
    author = "Wegmann, Anna  and
      Broek, Tijs A. Van Den  and
      Nguyen, Dong",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-main.52",
    pages = "882--912",
    abstract = "Best practices for high conflict conversations like counseling or customer support almost always include recommendations to paraphrase the previous speaker. Although paraphrase classification has received widespread attention in NLP, paraphrases are usually considered independent from context, and common models and datasets are not applicable to dialog settings. In this work, we investigate paraphrases across turns in dialog (e.g., Speaker 1: {``}That book is mine.{''} becomes Speaker 2: {``}That book is yours.{''}). We provide an operationalization of context-dependent paraphrases, and develop a training for crowd-workers to classify paraphrases in dialog. We introduce ContextDeP, a dataset with utterance pairs from NPR and CNN news interviews annotated for context-dependent paraphrases. To enable analyses on label variation, the dataset contains 5,581 annotations on 600 utterance pairs. We present promising results with in-context learning and with token classification models for automatic paraphrase detection in dialog.",
}