app.py

import gradio as gr
from transformers import pipeline, AutoTokenizer

# Define the model name
MODEL_NAME = "impresso-project/ner-stacked-bert-multilingual"

# Load the tokenizer and model using the pipeline
ner_tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

ner_pipeline = pipeline(
    "generic-ner",
    model=MODEL_NAME,
    tokenizer=ner_tokenizer,
    trust_remote_code=True,
    device="cpu",
)


# Helper function to align entities correctly and debug tokenization
def prepare_entities_for_highlight(text, results):
    entities = []
    seen_spans = set()  # Track the spans we have already added to avoid overlaps

    # Print debug info about tokenization
    print(f"Original text: {text}")

    # it should look like:
    # [{'entity': 'org.ent.pressagency.Reuters', 'score': np.float32(98.47), 'index': 78, 'word': 'Reuters', 'start': 440, 'end': 447}]
    for category, entity_list in results.items():
        for entity in entity_list:
            entity_span = (entity["start"], entity["end"])

            # Only add non-overlapping entities
            if entity_span not in seen_spans:
                seen_spans.add(entity_span)
                entity_text = text[
                    entity["start"] : entity["end"]
                ].strip()  # Ensure we're working with the correct portion of the text
                print(
                    f"Entity text: {entity_text}, Start: {entity['start']}, End: {entity['end']}, Type: {entity['entity']}"
                )
                entities.append(
                    {
                        "text": entity_text,
                        "score": entity["score"],
                        "start": entity["start"],
                        "end": entity["end"],
                        "label": entity["entity"],
                    }
                )

    # Sort entities by their start position
    entities = sorted(entities, key=lambda x: x["start"])

    return {"text": text, "entities": entities}


# Function to process the sentence and extract entities
def extract_entities(sentence):
    results = ner_pipeline(sentence)

    # Debugging the result format
    print(f"NER results: {results}")

    # Format the results for HighlightedText
    return prepare_entities_for_highlight(sentence, results)


# Create Gradio interface
def ner_app_interface():
    input_sentence = gr.Textbox(
        lines=5, label="Input Sentence", placeholder="Enter a sentence for NER..."
    )
    output_entities = gr.HighlightedText(label="Extracted Entities")

    # Interface definition
    interface = gr.Interface(
        fn=extract_entities,
        inputs=input_sentence,
        outputs=output_entities,
        title="Named Entity Recognition",
        description="Enter a sentence to extract named entities using the NER model from the Impresso project.",
        examples=[
            [
                "In the year 1789, King Louis XVI, ruler of France, convened the Estates-General at the Palace of Versailles."
            ]
        ],
        live=False,
    )

    interface.launch(share=True)


# Run the app
if __name__ == "__main__":
    ner_app_interface()