app.py

import gradio as gr
from transformers import pipeline

# Load the token classification pipeline using your model.
# Note: aggregation_strategy="simple" will merge sub-word tokens so that you get a single label per entity.
token_classifier = pipeline(
    "token-classification", 
    model="almo762/biobert-dependency-parsing-v1.7",
    aggregation_strategy="simple"  
)

def classify_tokens(text: str):
    # Run the model on the input text
    predictions = token_classifier(text)
    
    # Prepare a table output: each row contains Token, Label, Confidence Score
    results = []
    for pred in predictions:
        # The pipeline returns dictionaries with keys like "word", "entity_group", and "score".
        token = pred.get("word", "")
        label = pred.get("entity_group", "")
        score = pred.get("score", 0.0)
        results.append([token, label, f"{score:.2f}"])
    
    # If no predictions were found, return a friendly message
    if not results:
        return [["No tokens found", "", ""]]
    return results

examples = [
    "Mice transgenic for the human T cell leukemia virus (HTLV-I) Tax gene develop fibroblastic tumors that express NF-kappa B-inducible early genes.",
    "In vitro inhibition of NF-kappa B expression by antisense oligodeoxynucleotides (ODNs) inhibited growth of these culture-adapted Tax-transformed fibroblasts as well as an HTLV-I-transformed human lymphocyte line.",
    "In contrast, antisense inhibition of Tax itself had no apparent effect on cell growth.",
    "Mice treated with antisense to NF-kappa B ODNs showed rapid regression of transplanted fibrosarcomas.",
    "This suggests that NF-kappa B expression may be necessary for the maintenance of the malignant phenotype and provides a therapeutic approach for HTLV-I-associated disease."
]

# Build the Gradio interface.
demo = gr.Interface(
    fn=classify_tokens,
    inputs=gr.Textbox(lines=4, placeholder="Enter text here..."),
    outputs=gr.Dataframe(headers=["Token", "Label", "Confidence Score"]),
    title="BioBERT Dependency Parsing Demo",
    description="Enter text to see the token classification output: each token is shown with its assigned label and confidence score. The label can be interpreted after reading the paper: Viable Dependency Parsing as Sequence Labeling. The encode label we used here is Relative Part-of-Speech-based, since Part-of-Speech is easier to read than other types of encoding.",
    examples=examples
)

if __name__ == "__main__":
    demo.launch()