Topic Recognition, Sentiment Analysis, and Auto Reply in One Click
", unsafe_allow_html=True) # ------- Load ML Pipelines ------- @st.cache_resource def load_pipelines(): """ Load all required machine learning models for review analysis: - Topic classifier using zero-shot classification - Sentiment analysis model - Customer reply generation model Returns models and topic labels for use in the application """ # Define topic categories for classification topic_labels = [ "billing", "account access", "customer service", "loans", "fraud", "technical issue", "credit card", "mobile app", "branch service", "transaction delay", "account closure", "information error" ] # Ensure compatibility with CPU-only environments dtype = torch.float32 # Use float32 for better CPU compatibility # Load topic classification model topic_classifier = pipeline( "zero-shot-classification", model="MoritzLaurer/deberta-v3-base-zeroshot-v1", ) # Load sentiment analysis model sentiment_classifier = pipeline( "sentiment-analysis", model="cardiffnlp/twitter-roberta-base-sentiment-latest", ) # Load reply generation model model_name = "Leo66277/finetuned-tinyllama-customer-replies" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name) def generate_reply(text): """Generate a polite customer service reply based on the given review text""" prompt_text = f"Please write a short, polite English customer service reply to the following customer comment:\n{text}" inputs = tokenizer(prompt_text, return_tensors="pt", truncation=True, max_length=512) # Generate response with beam search for better quality with torch.no_grad(): gen_ids = model.generate( inputs.input_ids, max_length=inputs.input_ids.shape[1] + 120, pad_token_id=tokenizer.eos_token_id if tokenizer.eos_token_id else tokenizer.pad_token_id, num_beams=3, no_repeat_ngram_size=2, early_stopping=True ) # Clean up the generated text reply = tokenizer.decode(gen_ids[0][inputs.input_ids.shape[1]:], skip_special_tokens=True).strip() reply = reply.strip('"').replace('\n', ' ').replace(' ', ' ') return reply return topic_classifier, sentiment_classifier, generate_reply, topic_labels # ------- Page Layout -------- st.markdown("### Enter a review for instant analysis") # Updated example review as requested example_review = "BOA states on their website that holds are 2-7 days. I made a deposit, and the receipt states funds would be available in 2 days. Now, 13 days later, I am still waiting on my funds, and BOA can't give me a straight answer." # Text input area for user reviews user_review = st.text_area( "Please enter or paste a review below:", value=example_review, height=120 ) # Button to trigger analysis if st.button("Analyze"): if not user_review.strip(): # Validate user input st.warning("Please enter a valid review!") else: # Show user-friendly loading message with st.spinner("Analyzing your review..."): # Load models on first use to avoid loading message on startup if "topic_pipe" not in st.session_state: st.session_state.topic_pipe, st.session_state.sentiment_pipe, st.session_state.reply_generator, st.session_state.topic_labels = load_pipelines() # Run topic classification topic_result = st.session_state.topic_pipe(user_review, st.session_state.topic_labels, multi_label=False) topic = topic_result['labels'][0] # Removed confidence percentage as requested # Run sentiment analysis sentiment_result = st.session_state.sentiment_pipe(user_review) sentiment = sentiment_result[0]['label'] # Removed confidence percentage as requested # Generate auto reply reply_text = st.session_state.reply_generator(user_review) # Display results in a visually appealing format col1, col2 = st.columns(2) with col1: st.markdown(f"Topic:
{topic}Sentiment:
{sentiment}Auto-reply Suggestion:
{reply_text}