International Conference on Technological Intelligence and Business Strategies

Authors - Swe Swe Htun, Aye Nyein Mon
Abstract - Text classification has become a crucial task in natural language processing, especially for low-resourced languages in which limited annotated data and linguistic resources remain main challenges. This work presents inductive graph-based approach, GraphSAGE (Graph Sample and Aggregate), for text classification applying different word embedding models for Myanmar News classification. Experiments are conducted on eight Myanmar News categories (Business, Crime, Culture&Tourism, Educa-tion&Technology, Entertainment, Health, Politics, and Sports). The experiments show the effectiveness of BiLSTM (Bidirectional Long Short-Term Memory) and GraphSAGE architectures integrated with traditional and contextual embedding meth-ods, including TF-IDF (Term Frequency–Inverse Document Frequency), MyanBERTa, and mBERT (Multilingual Bidirectional Encoder Representations from Transformers). In the proposed work, transformer-based embeddings from pre-trained language models are extracted and combined with graph neural networks to capture both semantic and structural relationships among documents. A similarity graph is built by utilizing cosine similarity and k-nearest neighbor methods, and GraphSAGE is used to aggregate neighborhood information for inductive learning. The performance of graph-based models is compared to sequential deep learning technique based on BiLSTM. Experi-mental results reveal that graph-based approaches achieve better performance than BiLSTM-based models in all embedding settings. Among the evaluated models, mBERT with GraphSAGE gets the highest classification accuracy of 63%, followed by MyanBERTa with GraphSAGE with 60%. In contrast, MyanBERTa with BiLSTM and mBERT with BiLSTM yield 47% and 52% accuracy, respectively, whereas TF-IDF with GraphSAGE obtains 57% accuracy. The findings show that combining the contex-tual transformer embeddings with graph neural networks substantially enhance text classification performance by efficiently modeling semantic and relational information.