💬 NLP

GHTM: A Graph-based Hybrid Topic Modeling Approach with a Benchmark Dataset for the Low-Resource Bengali Language

This paper addresses the scarcity of resources for Bengali topic modeling by introducing GHTM, a novel graph-based hybrid architecture that combines TF-IDF-weighted GloVe embeddings, Graph Convolutional Networks, and Non-negative Matrix Factorization to achieve superior topic coherence and diversity, alongside the release of NCTBText, a diverse benchmark dataset of 8,650 textbook documents.

Original authors: Farhana Haque, Md. Abdur Rahman, Sumon Ahmed

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Farhana Haque, Md. Abdur Rahman, Sumon Ahmed

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a massive, chaotic library containing millions of books written in Bengali. The books are piled up randomly, there are no labels on the shelves, and you have no idea what any of them are about. Your goal is to organize this library into neat sections (like "Sports," "Science," "History") so people can find what they need.

This is the problem of Topic Modeling. It's a computer technique that reads thousands of documents and automatically figures out the main themes without a human having to read every single page.

For a long time, this was easy for English libraries but very hard for Bengali ones. Why? Because researchers didn't have good "maps" (datasets) or "tools" (algorithms) specifically designed for the Bengali language.

Here is a simple breakdown of what this paper does, using some everyday analogies.

1. The Problem: The "Messy Bengali Library"

Until now, trying to organize Bengali texts was like trying to sort a library using a map designed for English.

The Tools Were Old: Most researchers were using old, simple tools (like counting how many times a word appears) that missed the meaning of words.
The Data Was Boring: Almost all the data available was just news articles. It's like trying to understand a whole language only by reading the daily newspaper. You miss out on science, literature, history, and school textbooks.
No Standard Test: There was no "final exam" to see which tool was actually the best. Everyone used different test questions, so you couldn't compare them fairly.

2. The New Tool: GHTM (The "Smart Librarian")

The authors built a new, super-smart librarian robot called GHTM (Graph-based Hybrid Topic Model). Instead of just one trick, it uses a "hybrid" approach, combining the best parts of three different strategies:

Step 1: The "Highlighter" (TF-IDF + GloVe):
Imagine reading a book and highlighting the most important words. But instead of just highlighting, the robot also understands that "car" and "automobile" mean the same thing. It combines a statistical method (counting importance) with a semantic method (understanding meaning).
Step 2: The "Social Network" (Graph & GCN):
The robot looks at the books and asks, "Who are your friends?" It builds a social network where similar books are connected. Then, it uses a "Graph Convolutional Network" (GCN)—think of it as a gossip machine. If Book A talks to Book B, and Book B talks to Book C, the robot learns that A, B, and C are all part of the same "club" or topic. This helps it group things much better than just looking at words in isolation.
Step 3: The "Organizer" (NMF):
Finally, it takes all these connected groups and neatly sorts them into labeled boxes (topics) using a mathematical technique called Non-negative Matrix Factorization. It's like taking a pile of mixed Lego bricks and snapping them together into distinct, recognizable shapes.

The Result: GHTM is faster, smarter, and more accurate than the old tools. It found topics with much higher clarity (coherence) and variety (diversity) than any previous method.

3. The New Map: NCTBText (The "Textbook Treasure")

The authors realized that relying only on news was like trying to learn a language only from headlines. So, they created a brand new dataset called NCTBText.

Where did it come from? They scanned 8,650 pages from actual school textbooks in Bangladesh (covering subjects like Religion, Science, Agriculture, and ICT).
Why is it special? News talks about politics and crime. Textbooks talk about enzymes, history, and grammar. This dataset gives the computer a much richer vocabulary to learn from. It's like upgrading from a dictionary of slang to a full encyclopedia.
The Gift: They made this dataset public so other researchers can use it as a standard "test" for future tools.

4. The Big Test: Can it speak English too?

The authors didn't just test their robot on Bengali. They threw it into an English library (the famous "20Newsgroups" dataset) to see if it was a one-trick pony or a universal genius.

The Surprise: Even though GHTM was built for Bengali, it performed better than the best English-specific tools on the English test. This proves that the robot's logic is so sound that it works across languages. It's like building a car engine in Bangladesh that turns out to be faster than the best engines made in Germany.

5. Why Does This Matter?

For Bengali Speakers: It means we finally have a powerful, free tool to organize, search, and understand our own language's digital content.
For the World: It shows that you don't always need the most complex, expensive AI to get great results. Sometimes, mixing simple, proven methods (like counting words) with smart connections (graphs) works best.
For the Future: Because they shared their code and their textbook dataset, other researchers can now build on this work instead of starting from scratch.

In a Nutshell:
The authors took a messy, under-resourced problem (Bengali text analysis), built a smarter, hybrid robot (GHTM) to solve it, created a brand new library of textbooks (NCTBText) to train it, and proved that their robot is so good it can even organize English books better than the experts. They handed the keys to the whole world for free.

1. Problem Statement

Despite Bengali being the seventh most spoken language globally, topic modeling research for Bengali remains significantly underdeveloped compared to English. The field faces four critical challenges:

Lack of Systematic Benchmarking: There is no comprehensive comparison of traditional versus modern (neural, graph-based, cluster-based) topic models on Bengali data. Existing studies rely heavily on Latent Dirichlet Allocation (LDA) without evaluating state-of-the-art architectures.
Absence of Standardized Datasets: Most available Bengali datasets are scraped from newspapers, leading to a bias toward current events (politics, crime, sports) and a lack of topical diversity. There is no "gold standard" dataset comparable to the English 20Newsgroups.
Reproducibility Issues: Many proposed Bengali-specific models lack public code repositories, hindering validation and further research.
Resource Constraints: Existing advanced models (neural and graph-based) often suffer from high computational costs and poor scalability on large datasets, while traditional models fail to capture semantic relationships effectively.

2. Methodology: The GHTM Architecture

The authors propose GHTM (Graph-based Hybrid Topic Model), a novel architecture that strategically integrates statistical weighting, semantic embeddings, graph neural networks, and matrix factorization. The pipeline consists of four stages:

A. Text Vectorization (Hybrid Embeddings)

Instead of relying solely on Bag-of-Words (BoW) or raw embeddings, GHTM creates TF-IDF-weighted GloVe embeddings.

Process: It computes TF-IDF weights to capture term importance and multiplies them with pre-trained GloVe word vectors.
Formula: $X = S \cdot G$ , where $S$ is the sparse TF-IDF matrix and $G$ is the GloVe embedding matrix.
Benefit: This preserves both the statistical relevance of words within a document and the semantic relationships learned from global co-occurrence patterns.

B. Graph-based Representation Learning

To capture structural relationships between documents, GHTM constructs a K-Nearest Neighbor (KNN) graph where documents are nodes and edges represent semantic similarity.

Refinement via GCN: The model employs Graph Convolutional Networks (GCN), specifically using a Cluster-GCN architecture to handle large-scale memory constraints.
Training Objective: A hybrid loss function is used:
1. Hinge Loss: Preserves local structure by ensuring connected (similar) documents remain close in the embedding space.
2. Contrastive Loss (NT-Xent): Ensures global discriminability by pushing dissimilar documents apart.
Output: The GCN refines the initial embeddings into structurally informed dense vectors ( $Z$ ) that form coherent clusters representing latent topics.

C. Matrix Factorization

The refined embeddings are decomposed to extract interpretable topics.

Transformation: Since GCN outputs can contain negative values, an absolute value transformation ( $Z_{non-neg} = |Z|$ ) is applied to satisfy Non-negative Matrix Factorization (NMF) constraints.
Factorization: NMF decomposes the matrix into $W$ (document-topic distribution) and $H$ (topic representation).
Advantage: Unlike traditional NMF applied to sparse TF-IDF matrices, this operates on dense, graph-enriched embeddings, reducing lexical sparsity issues.

D. Topic Representation

Keyword Extraction: For each topic, the top $R$ representative documents are identified from matrix $W$ . The TF-IDF scores of words across these documents are aggregated to select the top $T$ keywords that define the topic.

3. Key Contributions

A. The GHTM Model

A novel hybrid architecture that outperforms traditional, neural, cluster-based, and graph-based baselines.
It balances computational efficiency with high topic coherence, avoiding the extreme resource demands of pure neural/graph models.

B. The NCTBText Dataset

Source: Curated from 8,650 text documents extracted from textbooks provided by the National Curriculum and Textbook Board (NCTB) of Bangladesh.
Diversity: Covers 8 distinct subject areas (Religion, Science, ICT, Agriculture, Business, etc.), breaking the "newspaper-centric" bias of existing Bengali corpora.
Validation: The dataset adheres to Zipf's Law and exhibits high lexical diversity (measured by TTR, MTLD, MATTR), making it a robust benchmark for future research.

C. Comprehensive Benchmarking

The first systematic evaluation of topic models across three Bengali datasets (Jamuna News, BanFakeNews, NCTBText) and one English benchmark (20Newsgroups).
Includes a wide range of baselines: LDA, LSA, NMF, Neural Topic Models (ProdLDA, ETM, CombinedTM), Cluster-based (Top2Vec, BERTopic), and Graph-based (GNTM, GCTM, Graph2Topic).

4. Experimental Results

Performance on Bengali Datasets

Coherence: GHTM achieved superior NPMI scores (0.27–0.28) and CV scores (0.82–0.90), outperforming the next best model (Top2Vec) by 8–17%.
Diversity: Achieved near-perfect Topic Diversity (TD: 0.96–1.00) and Inverted Rank-Biased Overlap (IRBO: 0.99–1.00).
Efficiency: Despite high performance, GHTM maintained low runtimes (13.57s – 231.15s), significantly faster than neural and graph-based baselines which often took thousands of seconds.
Generalization: GHTM performed consistently well across both news (Jamuna, BanFakeNews) and non-news (NCTBText) datasets, demonstrating robustness to domain shifts.

Cross-Lingual Generalization

Evaluated on the English 20Newsgroups dataset.
GHTM achieved the highest average NPMI (0.168) among graph-based models, outperforming established models like GINopic (0.126) and Graph2Topic (0.161), proving its language-agnostic capabilities.

Statistical Significance

Friedman Test: Confirmed statistically significant differences ( $p < 0.05$ ) between model performances.
Nemenyi Post-hoc Test: GHTM achieved the best average rank (1.0), significantly outperforming NeuralLDA and showing a strong trend over other baselines.

Qualitative Analysis

Human evaluation of topic words (e.g., in the "Sports" and "Electronics" categories) showed GHTM generated semantically coherent, domain-relevant keywords with zero out-of-topic words, whereas other models frequently included irrelevant terms or failed to capture specific domains.

5. Significance and Impact

Redefining the State-of-the-Art: GHTM establishes a new performance benchmark for Bengali topic modeling, proving that a hybrid approach combining statistical, semantic, and structural methods is superior to purely neural or purely algebraic approaches.
Resource Creation: The release of NCTBText fills a critical gap by providing a diverse, non-news benchmark, encouraging research into educational and academic text processing in Bengali.
Reproducibility: The authors have made the GHTM code and the NCTBText dataset publicly available, addressing the reproducibility crisis in low-resource NLP.
Low-Resource Applicability: The study demonstrates that for low-resource languages, carefully synthesizing complementary techniques (TF-IDF + Embeddings + GCN + NMF) can yield better results than simply scaling up complex neural architectures, offering a practical blueprint for other under-resourced languages.

Conclusion: The paper successfully addresses the lack of robust topic modeling frameworks for Bengali by introducing a high-performance, efficient hybrid model (GHTM) and a diverse, standardized dataset (NCTBText), setting the stage for future advancements in Bengali NLP.