BanglAssist: A Bengali-English Generative AI Chatbot for Code-Switching and Dialect-Handling in Customer Service

2025-03-28

AI Summary

This work addresses the limited capability of large language models (LLMs) in low-resource multilingual settings—specifically for Bangla–English code-mixing and dialectal variation—within customer service contexts in Bangladesh. We propose a lightweight, synergistic RAG-plus-domain-specific prompt engineering framework tailored to this scenario. Our approach innovatively integrates bilingual semantic alignment for retrieval, dialect-aware fine-tuning data construction, and culturally grounded prompt design, thereby enhancing cultural adaptability and cross-code-switching consistency. Evaluated on a real-world customer service test set, our method achieves 86.4% intent classification accuracy, reduces hallucination by 79.1%, and improves cross-code-switching response consistency by 32.7% over baseline LLMs. The framework delivers a scalable, robust technical paradigm for mixed-language dialogue systems in low-resource language settings.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Conversational AI/Dialog Systems
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

In recent years, large language models (LLMs) have demonstrated exponential improvements that promise transformative opportunities across various industries. Their ability to generate human-like text and ensure continuous availability facilitates the creation of interactive service chatbots aimed at enhancing customer experience and streamlining enterprise operations. Despite their potential, LLMs face critical challenges, such as a susceptibility to hallucinations and difficulties handling complex linguistic scenarios, notably code switching and dialectal variations. To address these challenges, this paper describes the design of a multilingual chatbot for Bengali-English customer service interactions utilizing retrieval-augmented generation (RAG) and targeted prompt engineering. This research provides valuable insights for the human-computer interaction (HCI) community, emphasizing the importance of designing systems that accommodate linguistic diversity to benefit both customers and businesses. By addressing the intersection of generative AI and cultural heterogeneity, this late-breaking work inspires future innovations in multilingual and multicultural HCI.

Problem

• Addressing Bengali-English code-switching in chatbots
• Handling dialect variations in customer service AI
• Reducing hallucinations in multilingual LLM applications

Innovation

• Uses retrieval-augmented generation for Bengali-English chatbot
• Implements targeted prompt engineering to reduce hallucinations
• Handles code-switching and dialect variations in customer service