🧠 YouTube Topic Modeling
(Phân tích chủ đề mô tả video YouTube bằng LDA – BERTopic – Top2Vec – CombinedTM)

---

🌟 1. Introduction (Giới thiệu)

This project analyzes Vietnamese YouTube video descriptions to discover latent semantic topics using modern topic modeling algorithms:
LDA, BERTopic, Top2Vec, and a hybrid CombinedTM (LDA + BERTopic).
(Dự án phân tích mô tả video YouTube tiếng Việt bằng LDA, BERTopic, Top2Vec và mô hình lai CombinedTM.)

Why this matters?

• YouTube descriptions reflect content trends & user behavior
• Topic modeling helps understand large-scale unstructured data
• Vietnamese text is noisy → requires deep preprocessing
• Useful for recommendation systems, content moderation, trend analysis

Dataset:

• 7,653 cleaned Vietnamese video descriptions
• Collected via YouTube Data API v3
• Includes title, tags, category_id, description, engagement metrics

---

🧼 2. Preprocessing (Tiền xử lý)

Applied steps:

• Lowercase text
• Remove emoji, URLs, emails, hashtags
• Remove advertisement keywords (“đăng ký”, “like”, “share”)
• Normalize Vietnamese characters
• Tokenize using Underthesea
• Remove stopwords
• Filter out short descriptions (< 5 tokens)

Outputs:

• clean_description
• tokenized list for LDA
• corpus + dictionary

---

🧠 3. Topic Modeling Methods (Các mô hình được sử dụng)

🔹 1. LDA (Latent Dirichlet Allocation)

(Mô hình thống kê phân phối xác suất chủ đề)

• Input: Bag-of-Words
• Optimized number of topics: 5
• Best Coherence: 0.549

---

🔹 2. BERTopic

(Embedding + UMAP + HDBSCAN + cTF-IDF)

• Multilingual Transformer
• Best Coherence: 0.714
• Best semantic quality across models
• Produces clear topic boundaries

---

🔹 3. Top2Vec

(Joint embedding for documents–words–topics)

• Automatically finds number of topics
• High Topic Diversity
• Lowest Coherence among 3 main models

---

🔹 4. CombinedTM (LDA + BERTopic)

(Mô hình lai kết hợp phân bố chủ đề LDA + embedding BERTopic)

Results:

• Coherence 0.663 (highest overall)
• Stable & interpretable
• Strong semantic grouping

---

📊 4. Results & Model Comparison (Kết quả và so sánh mô hình)

⭐ Overall Best Model: CombinedTM (LDA + BERTopic)

High balance of: ✔ Coherence
✔ Stability
✔ Purity
✔ Topic Diversity

---

🏆 5. Comparison Table (Bảng so sánh thuật toán)

📌 Table: Key Metrics Overview

Model	C_v	NPMI	NMI	ARI	Diversity	Nhận xét
LDA	0.549	0.039	0.193	0.092	0.800	Dễ diễn giải, ranh giới yếu
BERTopic	0.660	0.077	0.339	0.176	0.741	Ngữ nghĩa mạnh nhất, phân tách rõ
Top2Vec	0.421	0.000	0.148	0.043	0.800	Đa dạng cao, ngữ nghĩa yếu
CombinedTM	0.663	0.076	0.339	0.068	0.748	Tốt nhất tổng thể, cân bằng

---

🌀 6. Visualizations (Trực quan hóa)

Repository includes:

• WordCloud
• Bubble Chart
• Topic Cluster Maps
• Coherence plots
• Embedding visualizations (UMAP)

---

🧪 7. Key Findings (Kết luận quan trọng)

• BERTopic → strongest semantic model
• LDA → interpretable but overlapping
• Top2Vec → diverse but weak coherence
• CombinedTM → best hybrid model, highest Coherence
• Vietnamese descriptions require strong preprocessing
• Short-text topic modeling is challenging but solvable with embeddings

---

🚀 8. Applications (Ứng dụng thực tế)

• YouTube content trend analysis
• Automatic content tagging
• Detect emerging hot topics
• Improving recommendation algorithms
• Digital media research

---

👨‍🏫 9. Team (Nhóm thực hiện)

• Hà Thế Anh
• Nguyễn Nhật Nam
• Hoàng Quang Minh

Supervisor:

• Lê Nhật Tùng – HUTECH University