[3Xin2] Poster session 1

[3Xin2-78] Utterance control through QLORA tuning based on dialogue acts in a large language model using a Japanese dialogue corpus

〇Eiichi Shimizu¹, Masayasu Atsumi¹ (1.SOKA University)

Keywords:dialogue system, large language model, deep Learning

This paper describes a method for controlling system utterances based on dialog acts, which represent the intention of utterances, in chat dialog systems. To control system utterances based on dialogue acts, we apply QLORA tuning to a large language model using the Hazumi corpus, which is a Japanese dialogue corpus with dialogue act types, to generate appropriate system utterances with different prompts for different types of dialogue acts. Experiments were conducted to compare the effects of changing the prompts for each dialogue act type with those of applying the same prompts. The effect of QLORA tuning on utterance control was also evaluated through comparing models with and without tuning. The experimental results confirmed that QLORA tuning with different prompts for each dialogue act type was effective for utterance control and the effect was limited without QLORA tuning.