backend/punctuation.py

# Punctuation restoration — loads Oliver Guhr’s model and restores punctuation in raw text

import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

# Model 
MODEL_NAME = "oliverguhr/fullstop-punctuation-multilang-large"
DEVICE = 0 if torch.cuda.is_available() else -1

print(f"Loading punctuation model ({MODEL_NAME}) on {'GPU' if DEVICE == 0 else 'CPU'}...")

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForTokenClassification.from_pretrained(MODEL_NAME)

# pipeline for token classification
punctuation_pipeline = pipeline(
    "token-classification",
    model=model,
    tokenizer=tokenizer,
    device=DEVICE,
    aggregation_strategy="simple"
)

# Main function
def punctuate_text(text: str) -> str:
    """
    Restores punctuation in the given text using Oliver Guhr's model.
    Returns the punctuated text.
    """
    if not text.strip():
        return text

    try:
        results = punctuation_pipeline(text)

        punctuated_text = ""
        for item in results:
            word = item['word'].replace("▁", " ")  
            label = item['entity_group']

            # Map labels to punctuation marks
            if label == "COMMA":
                punctuated_text += word + ","
            elif label == "PERIOD":
                punctuated_text += word + "."
            elif label == "QUESTION":
                punctuated_text += word + "?"
            else:
                punctuated_text += word

        # Clean spacing
        return " ".join(punctuated_text.split())

    except Exception as e:
        print(f"[punctuate_text] Error: {e}")
        return text