JSAI2018

Presentation information

Oral presentation

Organized Session » [Organized Session] OS-13

[1H2-OS-13b] [Organized Session] OS-13

Tue. Jun 5, 2018 3:20 PM - 5:00 PM Room H (10F Sky Hall)

3:20 PM - 3:40 PM

[1H2-OS-13b-01] Evaluation of topic n-gram model for the utterance of AIWolf agent

〇Takashi OTSUKI1 (1. Yamagata University)

Keywords:AIWolf, topic n-gram model

AIWolf project presents "Are You a Werewolf?", the communication game with incomplete information, as a new standard game problem in the AI field. The project has also held competitions to realize the collective intelligence with participants. In recent competition, the champion agent uses the probabilistic model trained in advance. Therefore, it will be essential to utilize the more precise pre-trained models. In this article, we investigate the employment of topic n-gram model as the probabilistic model for the utterances generated by AIWolf agents. The evaluation of topic n-gram models of each role was carried out varying n from 1 to 4. It is shown that the perplexity of 2-gram model is far less than the one of 1-gram, and this concludes that we ought to use topic 2-gram model which brings good efficiency and robustness at the same time.