Generating Utterances for Companion Robots using Television Program Subtitles

Abstract

This study presents a method for generating utterances for companion robots that watch TV with people, using TV program subtitles. To enable the robot to automatically generate relevant utterances while watching TV, we created a dataset of approximately 12,000 utterances that were manually added to the collected TV subtitles. Using this dataset, we fine-tuned a large-scale language model to construct an utterance generation model. The proposed model generates utterances based on multiple keywords extracted from the subtitles as topics, while also taking into account the context of the subtitles by inputting them. The evaluation of the generated utterances revealed that approximately 88% of the sentences were natural Japanese, and approximately 75% were relevant and natural in the context of the TV program. Moreover, approximately 99% of the sentences contained the extracted keywords, indicating that our proposed method can generate diverse and contextually appropriate utterances containing the targeted topics. These findings provide evidence of the effectiveness of our approach in generating natural utterances for companion robots that watch TV with people.