Spaces:
Running
Running
| import os | |
| os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf-cache" | |
| os.environ["HF_HOME"] = "/tmp/hf-home" | |
| import nltk | |
| nltk.download("punkt", download_dir="/tmp/nltk_data") | |
| from sklearn.feature_extraction.text import TfidfVectorizer | |
| from sklearn.cluster import KMeans | |
| from sklearn.metrics.pairwise import cosine_similarity | |
| from nltk.tokenize import sent_tokenize | |
| from transformers import pipeline | |
| import numpy as np | |
| # Load summarizer and Q&A pipeline | |
| summarizer = pipeline("summarization", model="sshleifer/distilbart-cnn-12-6") | |
| qa_pipeline = pipeline("question-answering", model="distilbert-base-cased-distilled-squad") | |
| # --- Summarization Functions --- | |
| def summarize_review(text): | |
| """Standard transformer-based summarization""" | |
| return summarizer(text, max_length=max_len, min_length=min_len, do_sample=False)[0]["summary_text"] | |
| def smart_summarize(text, n_clusters=1): | |
| """ | |
| Clustering + cosine similarity-based summarization | |
| Selects most representative sentence(s) from each cluster | |
| """ | |
| sentences = sent_tokenize(text) | |
| if len(sentences) <= 1: | |
| return text | |
| tfidf = TfidfVectorizer(stop_words="english") | |
| tfidf_matrix = tfidf.fit_transform(sentences) | |
| if len(sentences) <= n_clusters: | |
| return " ".join(sentences) | |
| kmeans = KMeans(n_clusters=n_clusters, random_state=42).fit(tfidf_matrix) | |
| summary_sentences = [] | |
| for i in range(n_clusters): | |
| idx = np.where(kmeans.labels_ == i)[0] | |
| if not len(idx): | |
| continue | |
| avg_vector = np.asarray(tfidf_matrix[idx].mean(axis=0)) | |
| sim = cosine_similarity(avg_vector, tfidf_matrix[idx].toarray()) | |
| most_representative = sentences[idx[np.argmax(sim)]] | |
| summary_sentences.append(most_representative) | |
| return " ".join(sorted(summary_sentences, key=sentences.index)) | |
| # --- Rule-based Category Detectors --- | |
| def detect_industry(text): | |
| text = text.lower() | |
| if any(k in text for k in ["doctor", "hospital", "health", "pill", "med"]): | |
| return "Healthcare" | |
| if any(k in text for k in ["flight", "hotel", "trip", "booking"]): | |
| return "Travel" | |
| if any(k in text for k in ["bank", "loan", "credit", "payment"]): | |
| return "Banking" | |
| if any(k in text for k in ["gym", "trainer", "fitness", "workout"]): | |
| return "Fitness" | |
| if any(k in text for k in ["movie", "series", "stream", "video"]): | |
| return "Entertainment" | |
| if any(k in text for k in ["game", "gaming", "console"]): | |
| return "Gaming" | |
| if any(k in text for k in ["food", "delivery", "restaurant", "order"]): | |
| return "Food Delivery" | |
| if any(k in text for k in ["school", "university", "teacher", "course"]): | |
| return "Education" | |
| if any(k in text for k in ["insurance", "policy", "claim"]): | |
| return "Insurance" | |
| if any(k in text for k in ["property", "rent", "apartment", "house"]): | |
| return "Real Estate" | |
| if any(k in text for k in ["shop", "buy", "product", "phone", "amazon", "flipkart"]): | |
| return "E-commerce" | |
| return "Generic" | |
| def detect_product_category(text): | |
| text = text.lower() | |
| if any(k in text for k in ["mobile", "smartphone", "iphone", "samsung", "phone"]): | |
| return "Mobile Devices" | |
| if any(k in text for k in ["laptop", "macbook", "notebook", "chromebook"]): | |
| return "Laptops" | |
| if any(k in text for k in ["tv", "refrigerator", "microwave", "washer"]): | |
| return "Home Appliances" | |
| if any(k in text for k in ["watch", "band", "fitbit", "wearable"]): | |
| return "Wearables" | |
| if any(k in text for k in ["app", "portal", "site", "website"]): | |
| return "Web App" | |
| return "General" | |
| # --- Follow-up Q&A --- | |
| def answer_followup(text, question, verbosity="brief"): | |
| try: | |
| response = qa_pipeline({"question": question, "context": text}) | |
| answer = response.get("answer", "") | |
| if verbosity.lower() == "detailed": | |
| return f"Based on the review, the answer is: **{answer}**" | |
| return answer | |
| except Exception: | |
| return "Sorry, I couldn't generate a follow-up answer." | |