Enhancing BiGRU with a KAN Block for Legal Document Classification and Summarization

Researchers have developed a novel neural architecture combining Kolmogorov-Arnold Networks (KAN) with BiGRU models for classifying and summarizing legal documents in multilingual, low-resource settings. Tested on Bengali, English, and transliterated Bengali legal documents from Bangladesh, the hybrid model achieved 67.96% classification accuracy while demonstrating that KAN integration improved performance by over 10 percentage points.

This research addresses a critical gap in natural language processing: handling legal documents across multiple languages with limited training data. The integration of KAN modules into traditional BiGRU architectures represents an incremental advancement in sequence modeling, leveraging KAN's ability to learn non-linear relationships more efficiently than conventional neural components. The 10.62 percentage point improvement from ablation studies demonstrates KAN's practical value in this domain.

The work tackles genuine challenges in legal document processing: domain-specific terminology, multilingual complexity, long-range dependencies, and class imbalance. These obstacles plague real-world legal tech applications, particularly in South Asian jurisdictions where computational resources and annotated datasets remain scarce. The multilingual approach is especially relevant given Bangladesh's linguistic diversity.

From an industry perspective, improved document classification and summarization directly impact legal tech adoption. Automation of document review reduces costs for law firms and increases accessibility for smaller practices in developing markets. The research validates that specialized architectures can outperform general-purpose pretrained models when properly tuned for specific domains.

However, the reported metrics suggest meaningful limitations. A 67.96% classification accuracy and 0.38 ROUGE-1 score for summarization indicate the models remain imperfect for production deployment in high-stakes legal contexts where errors carry significant consequences. Future work should focus on scaling these techniques to larger datasets and exploring transfer learning approaches to improve performance in truly low-resource scenarios.

• →KAN-BiGRU hybrid architecture improved legal document classification accuracy from 57.34% to 67.96% through ablation testing.
• →The model handles multilingual legal text (Bengali, English, transliterated Bengali) in low-resource settings where labeled data is scarce.
• →Summarization performance (0.38 ROUGE-1) suggests the approach requires refinement before production-ready deployment in legal workflows.
• →Results demonstrate specialized neural architectures can outperform general pretrained language models on domain-specific NLP tasks.
• →The research addresses real market needs in South Asian legal tech where automation tools remain underdeveloped.