Title: Improving Legal Text Summarization and Classification via KAN-Integrated BiGRU

Abstract

This paper presents a new architecture that integrates a Kolmogorov-Arnold Network (KAN) block into a Bidirectional Gated Recurrent Unit (BiGRU) model, designed specifically for classifying and summarizing legal documents within a low-resource, multilingual environment. To address challenges such as domain-specific terminology, multilingual variations, long-range contextual dependencies, and imbalanced class distributions, the study utilizes a dataset sourced from Manupatra comprising legal texts in Bengali, English, and transliterated Bengali from Bangladesh.

The proposed classification framework combines a BiGRU with a KAN module, whereas the summarization component employs an attention-based GRU paired with a KAN head. Experimental results indicate that the classification model achieves an accuracy of 67.96% and an F1 score of 0.65. For summarization, the model attained ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.38, 0.23, and 0.31, respectively. An ablation study highlights the significant impact of the KAN integration, demonstrating that it boosts classification accuracy from 57.34% to 67.96%. Furthermore, the proposed method is evaluated against various baselines, including traditional machine learning algorithms and pretrained language models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC