Go to ACL ARR 2025 February homepageKAConv-Text: Kolmogorov-Arnold Convolutional Networks for Burmese Sentence Classification
This paper is the first study of the application of Kolmogorov-Arnold Convolutional Networks for Text data (KAConv-Text) on Burmese sentence classification across three tasks: hate speech detection (imbalanced binary), news classification (balanced multiclass), and ethnic language identification (imbalanced multiclass). Various embedding configurations were tested, utilizing random embeddings and fastText in both static and fine-tuned settings. Experiments were conducted with embedding dimensions of 100 and 300, comparing the CBOW and Skip-gram algorithms. Baseline models included Convolutional Neural Networks (CNN) and CNNs enhanced with a Kolmogorov-Arnold Network (KAN) classification layer (CNN-KAN). The proposed KAConv-Text with fine-tuned fastText embeddings achieved the best results, with 91.23% accuracy and a weighted F1-score of 0.9109 for hate speech detection, 92.66% accuracy and a weighted F1-score of 0.9267 for news classification, and 99.82% accuracy and a weighted F1-score of 0.9982 for ethnic language identification respectively.