Cross-lingual transfer learning with adapters for multilingual question generation subtitle

Abstract

In recent times, chatbots have become an integral component of online services, offering users a quick and efficient means to access information and accomplish tasks. However, creating a chatbot that can accurately comprehend and address user queries is a challenging task. To enhance chatbot performance, a promising technique is the use of question generation, which helps to align fresh user queries with existing solutions. This thesis project aims to develop an automatic question generation system that can produce relevant questions from knowledge bases, while also exploring the effectiveness of this approach in improving chatbot performance. In our research, we experiment with two distinct methods: one based on adapter training and the other leveraging the GPT-3.5 turbo model. Through various evaluations of the results generated by these systems, we find that training both language and task adapters is an effective approach for cross-lingual transfer learning, outperforming our baseline metrics. Taking into account computational resources and integration ease, we selected the GPT model for incorporation into the chatbot system, thereby enhancing its capabilities via question generation.