📰 Vietnamese-News-Topic-Modeling

---

📌 Overview (Tổng quan)

This project applies multiple Topic Modeling approaches—including LDA, NMF, LSA, and BERTopic—to identify and analyze latent topics from 7,052 preprocessed Vietnamese online news articles.
(Dự án áp dụng nhiều phương pháp Topic Modeling—LDA, NMF, LSA, BERTopic—để trích xuất & phân tích chủ đề tiềm ẩn từ 7.052 bài báo điện tử Việt Nam đã tiền xử lý.)

It includes:

• Full web scraping pipeline (Selenium + BeautifulSoup)
• Text cleaning, normalization, tokenization
• TF-IDF, BoW, and PhoBERT embeddings
• Training & evaluation across 4 topic models
• Coherence, Perplexity, Reconstruction Error, ARI, NMI evaluation
• Final comparison & conclusions

📄 Full content extracted from: :contentReference[oaicite:1]{index=1}

---

📚 Data Pipeline (Quy trình dữ liệu)

1. Data Collection (Thu thập dữ liệu)

• Automatic crawling using Selenium & BeautifulSoup
• Extracted fields: title, category, content, pub_date, author, link
• 7,580 articles collected → 7,052 final cleaned samples
• Saved as CSV for reproducibility

2. Preprocessing (Tiền xử lý)

Includes:

• Lowercasing
• Removing URLs, emojis, digits, and special characters
• Vietnamese word segmentation using Underthesea
• Removing duplicates using MinHash + LSH
• Removing short documents (<100 tokens)
• Merging rare categories
• Exporting cleaned dataset for modeling

3. Feature Engineering (Biểu diễn đặc trưng)

• TF-IDF → used for NMF & LSA
• Bag-of-Words → used for LDA
• PhoBERT / Vietnamese-SBERT embeddings (768-dim) → used for BERTopic

---

🔥 Models Used (Các mô hình sử dụng)

Category	Models
Classical Topic Models	LDA, NMF, LSA
Embedding-based Models	PhoBERT + BERTopic
Clustering Algorithms	HDBSCAN, KMeans
Evaluations	Coherence, Perplexity, Reconstruction Error, ARI, NMI

---

📊 Algorithm Comparison Table

(Bảng so sánh mô hình – đẹp & chuyên nghiệp)

Table 1 — Summary of Optimal Results

(Tóm tắt kết quả tốt nhất của từng mô hình)

Model	Optimal K	Coherence (c_v)	Other Metrics	Notes (Ghi chú)
NMF (TF-IDF)	13	0.8405	Rec. Error ↓ 80.24	Best coherence; balanced topics; easy interpretation
BERTopic (SBERT)	11	0.7324	Outliers: 600 docs	Strong semantic grouping; modern embeddings
LDA (BoW)	13	0.6642	—	Clean topics; probabilistic; stable
LSA + KMeans	12	0.4040	Var. Explained 0.0681	Weak separation; overlapping topics

---

🧠 Detailed Model Insights (Phân tích chi tiết từng mô hình)

1️⃣ LDA – Latent Dirichlet Allocation

• Works on Bag-of-Words
• Best at K = 13
• Coherence = 0.6642
• Produces clean, interpretable topics
• Topics include: education, health, economy, law, lifestyle, sports...

2️⃣ NMF – Non-negative Matrix Factorization

• Works on TF-IDF
• Best at K = 13
• Highest coherence: 0.8405
• Most stable and interpretable result
• Topics: ESG, family, auto, sports, health, Ukraine–Russia, entertainment...

3️⃣ LSA + KMeans

• Works on TF-IDF + SVD
• Best at K = 12
• Coherence = 0.4040 (low)
• Topics overlap heavily → not ideal for Vietnamese text

4️⃣ BERTopic (PhoBERT + UMAP + HDBSCAN)

• Embedding-based, no need to choose K
• Extracted 11 topics
• Coherence = 0.7324
• Strong semantic clustering
• Some outliers (~600 docs)

---

🧪 Evaluation Metrics (Các chỉ số đánh giá)

Intrinsic (Nội tại)

• Coherence (c_v): Topic interpretability
• Perplexity (LDA)
• Reconstruction Error (NMF)
• Explained Variance (LSA)

Extrinsic (Ngoại tại)

Model	ARI	NMI
NMF	0.4295	0.5417
LDA	0.3989	0.5234
LSA	0.3373	0.5300

➡️ NMF provides the best match to real categories.

---

🚀 Applications (Ứng dụng thực tế)

• News trend analysis (Phân tích xu hướng báo chí)
• Media monitoring (Giám sát truyền thông)
• Policy insights (Hỗ trợ hoạch định chính sách)
• Contextual advertising (Quảng cáo theo ngữ cảnh)
• Research & social science (Nghiên cứu học thuật)
• Building recommender systems (Hệ thống gợi ý bài viết)

---

⚠️ Limitations (Hạn chế)

• Single data source → limited generalization
• LSA coherence low
• BERTopic requires GPU & produces outlier cluster
• Classical models require manual K tuning

---

🔮 Future Work (Hướng phát triển)

• Expand dataset to multiple news sources
• Use advanced neural topic models: ProdLDA, ETM
• Improve Vietnamese embeddings (vBERT, PhoBERT-large)
• Deploy real-time topic explorer dashboard
• Incorporate human evaluation (expert labeling)

---

👨‍💻 Authors (Tác giả)

• Mai Thanh Phúc
• Hoàng Thị Yến Nhi
• Trần Trọng Thành
• GVHD: Lê Nhật Tùng

---

📚 Citation (Trích dẫn)

If you use this project in research, please cite:

Mai Thanh Phúc, Hoàng Thị Yến Nhi, Trần Trọng Thành, Lê Nhật Tùng. “Topic Modeling Approaches for Vietnamese Online News Analysis.” HUTECH University, 2025.