import streamlit as st import torch from transformers import AutoTokenizer from semviqa.ser.qatc_model import QATCForQuestionAnswering from semviqa.tvc.model import ClaimModelForClassification from semviqa.ser.ser_eval import extract_evidence_tfidf_qatc from semviqa.tvc.tvc_eval import classify_claim import time import pandas as pd import os import psutil import gc import numpy as np from functools import lru_cache import threading from concurrent.futures import ThreadPoolExecutor import torch.nn.functional as F # Set environment variables to optimize CPU performance os.environ["OMP_NUM_THREADS"] = str(psutil.cpu_count(logical=False)) os.environ["MKL_NUM_THREADS"] = str(psutil.cpu_count(logical=False)) torch.set_num_threads(psutil.cpu_count(logical=False)) # Set device globally DEVICE = "cuda" if torch.cuda.is_available() else "cpu" # Cache for model outputs @lru_cache(maxsize=1000) def cached_classify_claim(claim, evidence, model_name, is_bc=False): tokenizer, model = load_model(model_name, ClaimModelForClassification, is_bc=is_bc, device=DEVICE) with torch.no_grad(): prob, pred = classify_claim(claim, evidence, model, tokenizer, DEVICE) return prob, pred # Optimized model loading with caching @st.cache_resource(ttl=3600) # Cache for 1 hour def load_model(model_name, model_class, is_bc=False, device=None): if device is None: device = "cuda" if torch.cuda.is_available() else "cpu" tokenizer = AutoTokenizer.from_pretrained(model_name) model = model_class.from_pretrained(model_name, num_labels=3 if not is_bc else 2) model.eval() # Optimize model for inference if device == "cuda": model = model.half() # Use FP16 for faster inference torch.cuda.empty_cache() model.to(device) return tokenizer, model # Optimized text preprocessing @st.cache_data(ttl=3600) def preprocess_text(text): # Add any text cleaning or normalization here return text.strip() # Batch processing for evidence extraction def batch_extract_evidence(claims, contexts, model_qatc, tokenizer_qatc, batch_size=4): results = [] for i in range(0, len(claims), batch_size): batch_claims = claims[i:i + batch_size] batch_contexts = contexts[i:i + batch_size] with torch.no_grad(): batch_results = [ extract_evidence_tfidf_qatc( claim, context, model_qatc, tokenizer_qatc, DEVICE, confidence_threshold=0.5, length_ratio_threshold=0.5 ) for claim, context in zip(batch_claims, batch_contexts) ] results.extend(batch_results) return results # Optimized verification function with parallel processing def perform_verification(claim, context, model_qatc, tokenizer_qatc, model_tc, tokenizer_tc, model_bc, tokenizer_bc, tfidf_threshold, length_ratio_threshold): # Extract evidence with optimized settings evidence_start_time = time.time() evidence = extract_evidence_tfidf_qatc( claim, context, model_qatc, tokenizer_qatc, DEVICE, confidence_threshold=tfidf_threshold, length_ratio_threshold=length_ratio_threshold ) evidence_time = time.time() - evidence_start_time # Clear memory after evidence extraction if DEVICE == "cuda": torch.cuda.empty_cache() gc.collect() verdict_start_time = time.time() # Parallel classification using ThreadPoolExecutor with ThreadPoolExecutor(max_workers=2) as executor: future_tc = executor.submit(cached_classify_claim, claim, evidence, tc_model_name, False) future_bc = executor.submit(cached_classify_claim, claim, evidence, bc_model_name, True) prob3class, pred_tc = future_tc.result() prob2class, pred_bc = future_bc.result() with torch.no_grad(): verdict = "NEI" if pred_tc != 0: verdict = "SUPPORTED" if pred_bc == 0 else "REFUTED" if prob2class > prob3class else ["NEI", "SUPPORTED", "REFUTED"][pred_tc] verdict_time = time.time() - verdict_start_time return { "evidence": evidence, "verdict": verdict, "evidence_time": evidence_time, "verdict_time": verdict_time, "prob3class": prob3class.item() if isinstance(prob3class, torch.Tensor) else prob3class, "pred_tc": pred_tc, "prob2class": prob2class.item() if isinstance(prob2class, torch.Tensor) else prob2class, "pred_bc": pred_bc } # Add performance monitoring def monitor_performance(): if DEVICE == "cuda": return { "gpu_memory_used": torch.cuda.memory_allocated() / 1024**2, "gpu_memory_cached": torch.cuda.memory_reserved() / 1024**2, "cpu_percent": psutil.cpu_percent(), "memory_percent": psutil.virtual_memory().percent } return { "cpu_percent": psutil.cpu_percent(), "memory_percent": psutil.virtual_memory().percent } # Set page configuration st.set_page_config( page_title="SemViQA - A Semantic Question Answering System for Vietnamese Information Fact-Checking", page_icon="🔍", layout="wide", initial_sidebar_state="expanded" ) # Custom CSS st.markdown(""" """, unsafe_allow_html=True) # Main header st.markdown("""
SemViQA
A Semantic Question Answering System for Vietnamese Information Fact-Checking
""", unsafe_allow_html=True) # Sidebar with st.sidebar: st.markdown("### ⚙️ System Settings") # Model selection st.markdown("#### 🧠 Model Selection") qatc_model_name = st.selectbox( "QATC Model", [ "SemViQA/qatc-infoxlm-viwikifc", "SemViQA/qatc-infoxlm-isedsc01", "SemViQA/qatc-vimrc-viwikifc", "SemViQA/qatc-vimrc-isedsc01" ] ) bc_model_name = st.selectbox( "Binary Classification Model", [ "SemViQA/bc-xlmr-viwikifc", "SemViQA/bc-xlmr-isedsc01", "SemViQA/bc-infoxlm-viwikifc", "SemViQA/bc-infoxlm-isedsc01", "SemViQA/bc-erniem-viwikifc", "SemViQA/bc-erniem-isedsc01" ] ) tc_model_name = st.selectbox( "Three-Class Classification Model", [ "SemViQA/tc-xlmr-viwikifc", "SemViQA/tc-xlmr-isedsc01", "SemViQA/tc-infoxlm-viwikifc", "SemViQA/tc-infoxlm-isedsc01", "SemViQA/tc-erniem-viwikifc", "SemViQA/tc-erniem-isedsc01" ] ) # Threshold settings st.markdown("#### ⚖️ Analysis Thresholds") tfidf_threshold = st.slider( "Confidence Threshold", 0.0, 1.0, 0.5, help="Adjust sensitivity in evidence search" ) length_ratio_threshold = st.slider( "Length Ratio Threshold", 0.1, 1.0, 0.5, help="Adjust maximum evidence length" ) # Display settings st.markdown("#### 👁️ Display") show_details = st.checkbox( "Show Probability Details", value=False, help="Display detailed probability information" ) # Performance settings st.markdown("#### ⚡ Performance") num_threads = st.slider( "CPU Threads", 1, psutil.cpu_count(), psutil.cpu_count(logical=False), help="Adjust processing performance" ) os.environ["OMP_NUM_THREADS"] = str(num_threads) os.environ["MKL_NUM_THREADS"] = str(num_threads) # Main content tabs = st.tabs(["🔍 Verify", "📊 History", "ℹ️ Info"]) tokenizer_qatc, model_qatc = load_model(qatc_model_name, QATCForQuestionAnswering, device=DEVICE) tokenizer_bc, model_bc = load_model(bc_model_name, ClaimModelForClassification, is_bc=True, device=DEVICE) tokenizer_tc, model_tc = load_model(tc_model_name, ClaimModelForClassification, device=DEVICE) verdict_icons = { "SUPPORTED": "✅", "REFUTED": "❌", "NEI": "⚠️" } # --- Tab Verify --- with tabs[0]: col1, col2 = st.columns([2, 1]) with col1: st.markdown("### 📝 Input Information") claim = st.text_area( "Claim to Verify", "Chiến tranh với Campuchia đã kết thúc trước khi Việt Nam thống nhất.", help="Enter the claim to be verified" ) context = st.text_area( "Context", "Sau khi thống nhất, Việt Nam tiếp tục gặp khó khăn do sự sụp đổ và tan rã của đồng minh Liên Xô cùng Khối phía Đông, các lệnh cấm vận của Hoa Kỳ, chiến tranh với Campuchia, biên giới giáp Trung Quốc và hậu quả của chính sách bao cấp sau nhiều năm áp dụng. Năm 1986, Đảng Cộng sản ban hành cải cách đổi mới, tạo điều kiện hình thành kinh tế thị trường và hội nhập sâu rộng. Cải cách đổi mới kết hợp cùng quy mô dân số lớn đưa Việt Nam trở thành một trong những nước đang phát triển có tốc độ tăng trưởng thuộc nhóm nhanh nhất thế giới, được coi là Hổ mới châu Á dù cho vẫn gặp phải những thách thức như tham nhũng, tội phạm gia tăng, ô nhiễm môi trường và phúc lợi xã hội chưa đầy đủ. Ngoài ra, giới bất đồng chính kiến, chính phủ một số nước phương Tây và các tổ chức theo dõi nhân quyền có quan điểm chỉ trích hồ sơ nhân quyền của Việt Nam liên quan đến các vấn đề tôn giáo, kiểm duyệt truyền thông, hạn chế hoạt động ủng hộ nhân quyền cùng các quyền tự do dân sự.", help="Enter context or reference text" ) verify_button = st.button("🔍 Verify", use_container_width=True) with col2: st.markdown("### 📊 Results") if verify_button: with st.spinner("Verifying..."): # Preprocess texts preprocessed_claim = preprocess_text(claim) preprocessed_context = preprocess_text(context) # Clear memory and perform verification gc.collect() if DEVICE == "cuda": torch.cuda.empty_cache() start_time = time.time() # Monitor initial performance initial_perf = monitor_performance() result = perform_verification( preprocessed_claim, preprocessed_context, model_qatc, tokenizer_qatc, model_tc, tokenizer_tc, model_bc, tokenizer_bc, tfidf_threshold, length_ratio_threshold ) total_time = time.time() - start_time # Monitor final performance final_perf = monitor_performance() # Format details details = "" if show_details: gpu_memory_used = f"{float(final_perf.get('gpu_memory_used', 0)):.2f} MB" if DEVICE == "cuda" else "N/A" gpu_memory_cached = f"{float(final_perf.get('gpu_memory_cached', 0)):.2f} MB" if DEVICE == "cuda" else "N/A" details = f""" Details: - 3-Class Probability: {result['prob3class']:.2f} - 3-Class Predicted Label: {['NEI', 'SUPPORTED', 'REFUTED'][result['pred_tc']]} - 2-Class Probability: {result['prob2class']:.2f} 2-Class Predicted Label: {['SUPPORTED', 'REFUTED'][result['pred_bc']] if isinstance(result['pred_bc'], int) and result['pred_tc'] != 0 else 'Not used'} Performance Metrics: - GPU Memory Used: {gpu_memory_used} - GPU Memory Cached: {gpu_memory_cached} - CPU Usage: {final_perf['cpu_percent']}% - Memory Usage: {final_perf['memory_percent']}% """ # Store result with performance metrics st.session_state.latest_result = { "claim": claim, "evidence": result['evidence'], "verdict": result['verdict'], "evidence_time": result['evidence_time'], "verdict_time": result['verdict_time'], "total_time": total_time, "details": details, "qatc_model": qatc_model_name, "bc_model": bc_model_name, "tc_model": tc_model_name, "performance_metrics": final_perf } # Add to history if 'history' not in st.session_state: st.session_state.history = [] st.session_state.history.append(st.session_state.latest_result) # Display result with performance metrics res = st.session_state.latest_result verdict_class = { "SUPPORTED": "verdict-supported", "REFUTED": "verdict-refuted", "NEI": "verdict-nei" }.get(res['verdict'], "") gpu_memory_text = ( f"
  • GPU Memory: {float(res['performance_metrics'].get('gpu_memory_used', 0)):.2f} MB
    • " if DEVICE == "cuda" else "
  • GPU Memory: N/A
    • " ) st.markdown(f"""

    Verification Results

    Claim: {res['claim']}

    Evidence: {res['evidence']}

    {verdict_icons.get(res['verdict'], '')} {res['verdict']}

    Evidence Extraction Time: {res['evidence_time']:.2f} seconds

    Classification Time: {res['verdict_time']:.2f} seconds

    Total Time: {res['total_time']:.2f} seconds

    Performance:

    {f"
    {res['details']}
    " if show_details else ""}
    """, unsafe_allow_html=True) # Download button with performance metrics result_text = f""" Claim: {res['claim']} Evidence: {res['evidence']} Verdict: {res['verdict']} Details: {res['details']} Performance: - CPU: {res['performance_metrics']['cpu_percent']}% - RAM: {res['performance_metrics']['memory_percent']}% - GPU Memory: {f"{float(res['performance_metrics'].get('gpu_memory_used', 0)):.2f} MB" if DEVICE == "cuda" else "N/A"} """ st.download_button( "📥 Download Results", data=result_text, file_name="verification_results.txt", mime="text/plain" ) else: st.info("Please enter information and click Verify to begin.") # --- Tab History --- with tabs[1]: st.markdown("### 📊 Verification History") if 'history' in st.session_state and st.session_state.history: # Download full history history_df = pd.DataFrame(st.session_state.history) st.download_button( "📥 Download Full History", data=history_df.to_csv(index=False).encode('utf-8'), file_name="verification_history.csv", mime="text/csv" ) # Display history for idx, record in enumerate(reversed(st.session_state.history), 1): st.markdown(f"""

    Verification #{idx}

    Claim: {record['claim']}

    Verdict: {verdict_icons.get(record['verdict'], '')} {record['verdict']}

    Time: {record['total_time']:.2f} seconds

    """, unsafe_allow_html=True) else: st.info("No verification history available.") # --- Tab Info --- with tabs[2]: st.markdown(""" ### ℹ️ About SemViQA **Author:** [**Nam V. Nguyen**](https://github.com/DAVID-NGUYEN-S16), [**Dien X. Tran**](https://github.com/xndien2004), Thanh T. Tran, Anh T. Hoang, Tai V. Duong, Di T. Le, Phuc-Lu Le SemViQA is a cutting-edge Vietnamese fact-checking system designed to combat misinformation. It leverages semantic-based evidence retrieval (SER) and a two-step verdict classification (TVC) approach to verify claims efficiently. By combining TF-IDF with a Question Answering Token Classifier (QATC), SemViQA improves accuracy while reducing inference time. Achieving state-of-the-art performance, it has set new benchmarks on ViWikiFC (80.82% accuracy) and ISE-DSC01 (78.97% accuracy) datasets. With its 7x speed boost, SemViQA is a powerful tool for ensuring information integrity in the Vietnamese language. #### 🔍 How to Use 1. Enter the claim to verify 2. Enter context or reference text 3. Adjust parameters in Settings if needed 4. Click Verify button #### ⚙️ Parameters - **Confidence Threshold:** Adjust sensitivity in evidence search - **Length Ratio Threshold:** An important parameter in the evidence retrieval process. It determines how text segments are processed when compared to the length of the claim to be verified. - **CPU Threads:** Adjust processing performance #### 📊 Results - **SUPPORTED:** The claim is supported by evidence - **REFUTED:** The claim is refuted by evidence - **NEI:** Not enough information to conclude """)