| 講演抄録/キーワード |
| 講演名 |
2019-06-10 10:30
Impression Prediction of Oral Presentation Using LSTM with Dot-product Attention Mechanism ○Shengzhou Yi・Xueting Wang・Toshihiko Yamasaki(UTokyo) MVE2019-1 |
| 抄録 |
(和) |
(まだ登録されていません) |
| (英) |
For automatically evaluating oral presentation, we propose an end-to-end system to predict audience’s impression on speech video. Our framework is a multimodal neural network including two Long Short-Term Memory (LSTM) with dot-product attention mechanism to learn linguistic feature and acoustic feature respectively for our classification task, as well as a hidden network to consider the correlation between different types of feature representations for model-level fusion. We utilize 2,445 videos with official captions and users’ ratings from TED Talks. The experiment result shows the good performance of our proposal can recognize audience’s 14 types of impression with the average accuracy of 85.3%. |
| キーワード |
(和) |
/ / / / / / / |
| (英) |
Presentation Analysis / Multimodal Network / End-to-End System / TED Talks / / / / |
| 文献情報 |
信学技報, vol. 119, no. 75, MVE2019-1, pp. 1-6, 2019年6月. |
| 資料番号 |
MVE2019-1 |
| 発行日 |
2019-06-03 (MVE) |
| ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
| PDFダウンロード |
MVE2019-1 |