app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline
import torch
import torch.nn.functional as F

@st.cache_resource
def load_model():
    absa_tokenizer = AutoTokenizer.from_pretrained("yangheng/deberta-v3-base-absa-v1.1", use_fast=False)
    absa_model = AutoModelForSequenceClassification.from_pretrained("yangheng/deberta-v3-base-absa-v1.1")
    token_classifier = pipeline(
        model="thainq107/abte-restaurants-distilbert-base-uncased",
        aggregation_strategy="simple"
    )
    return absa_model, absa_tokenizer, token_classifier

absa_model, absa_tokenizer, token_classifier = load_model()

def inference(review):
    aspects = token_classifier(review)
    aspects = [aspect['word'] for aspect in aspects]
    results = {}
    for aspect in aspects:
        # Check if the aspect is mentioned in the review
        if aspect.lower() in review.lower():
            inputs = absa_tokenizer(f"[CLS] {review} [SEP] {aspect} [SEP]", return_tensors="pt")
            outputs = absa_model(**inputs)
            probs = F.softmax(outputs.logits, dim=1)
            probs = probs.detach().numpy()[0]

            # Extract the label with the highest probability
            max_label, max_prob = max(zip(["negative", "neutral", "positive"], probs), key=lambda x: x[1])

            results[aspect] = max_label

    return results

st.title("ABSA - Aspect-Based Sentiment Analysis")

text = st.text_area("Nhập câu cần phân tích:", "The battery life is great, but the screen is dim.")

if st.button("Phân tích cảm xúc"):
    results = inference(text)
    for aspect, label in results.items():
        st.markdown(f"{aspect} ➝ {label}")