Presentation | 2023-03-01 A Study of Utterance Generation Models based on TV Programs and Interests Yuta Hagio, Makoto Okuda, Marina Kamimura, Yutaka Kaneko, Hisayuki Ohmata, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | In this paper, we discuss the results of a study of an utterance generation model based on TV programs and specifiedinterest words, with the aim of implementing it in a companion robot that watches TV with humans. In order to automaticallygenerate the robot’s utterances while watching TV, we created a TV program-related utterance dataset by adding about 12,000robot utterances to the program’s captions. Using this dataset, we built an utterance generation model by fine-tuning a large-scalegeneral-purpose language model for chat dialog. Our model can generate utterances with the given specified interest wordas the topic, taking into account the context of the TV program using its captions and the interest word. As a result of evaluationof the sentences generated by our model, we confirmed that about 88% of the sentences are natural Japanese sentences, andabout 75% of the sentences are natural sentences in the context of the program. In addition, about 83% of the utterancesreflected the specified interest words. It was found that our model can generate utterances that are natural in the context of theprogram’s captions and that suggest interest in the specified interest words. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Companion robot / TV-viewing / Utterance generation / Caption data / Interest words |
Paper # | BioX2022-64,CNR2022-30 |
Date of Issue | 2023-02-22 (BioX, CNR) |
Conference Information | |
Committee | CNR / BioX |
---|---|
Conference Date | 2023/3/1(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | Masayuki Kanbara(NAIST) / Hitoshi Imaoka(NEC) |
Vice Chair | Yuri Nishikawa(AIST) / Norihiko Okui(KDDI Research) / Naoyuki Takada(SECOM) |
Secretary | Yuri Nishikawa(Toshiba) / Norihiko Okui(Nagoya Sangyo Univ.) / Naoyuki Takada(NEC) |
Assistant | Yuta Hoshi(NHK) / Junji Yamato(Kogakuin Univ.) / Hiroyuki Suzuki(Gunma Univ) / Shinichi Shirakawa(Yokohama National Univ.) |
Paper Information | |
Registration To | Technical Committee on Cloud Network Robotics / Technical Committee on Biometrics |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | A Study of Utterance Generation Models based on TV Programs and Interests |
Sub Title (in English) | |
Keyword(1) | Companion robot |
Keyword(2) | TV-viewing |
Keyword(3) | Utterance generation |
Keyword(4) | Caption data |
Keyword(5) | Interest words |
1st Author's Name | Yuta Hagio |
1st Author's Affiliation | Japan Broadcasting Corporation(NHK) |
2nd Author's Name | Makoto Okuda |
2nd Author's Affiliation | Japan Broadcasting Corporation(NHK) |
3rd Author's Name | Marina Kamimura |
3rd Author's Affiliation | Japan Broadcasting Corporation(NHK) |
4th Author's Name | Yutaka Kaneko |
4th Author's Affiliation | Japan Broadcasting Corporation(NHK) |
5th Author's Name | Hisayuki Ohmata |
5th Author's Affiliation | Japan Broadcasting Corporation(NHK) |
Date | 2023-03-01 |
Paper # | BioX2022-64,CNR2022-30 |
Volume (vol) | vol.122 |
Number (no) | BioX-394,CNR-395 |
Page | pp.pp.13-18(BioX), pp.13-18(CNR), |
#Pages | 6 |
Date of Issue | 2023-02-22 (BioX, CNR) |