On the Effectiveness of LLM in Question-Answering Systems

At Kindai University, the question-and-answer system on Slack faced challenges due to a high volume of similar inquiries related to class content. A question-answering assistance system using large language models (LLMs) trained on historical questions is introduced in this paper. The effectiveness of three existing Japanese LLMs using two methods: a full parameter update method and a technique called LoRa are investigated. The results show that the full parameter update method achieved lower cross-entropy loss, while the combination of both methods imploving responses.