Enhancing Conversational AI Model Performance and Explainability for Sinhala-English Bilingual Speakers

Abstract

Natural language processing has become essential to modern conversational tools and dialogue engines, including Chatbots. However, applying natural language processing to low-resource languages is challenging due to their lack of digital presence. Sinhala is the native language of approximately nineteen million people in Sri Lanka and is one of many low-resource languages. Moreover, the increase in using code-switching: alternating two or more languages within the same conversation, and code-mixing: the practice of representing words of a language using characters of another language, has become another major issue when processing natural languages. Apart from natural language processing, the explainability of opaque machine learning models utilized in chatbots has become another prominent concern. None of the existing modern chatbot development platforms supports explainability and relies on a performance score such as accuracy or f1-score. This paper proposes a no-code chatbot development platform with a series of built-in novel natural language processing, model evaluation, and explainability tools to tackle the problems of processing Sinhala-English code-switching and code-mixing natural language data and model evaluation in modern chatbot development platforms.