[3A1-GS-6] Language media processing

[3A1-GS-6-05] Knowledge selection focusing on standard/non-standard information in database

〇Haruki Hatakeyama¹, Ryouhei Kaneda¹, Daisuke Kawakubo¹, Masaki Shuzo¹, Eisaku Maeda¹ (1. Tokyo Denki University)

Keywords:Knowledge selection, RoBERTa, Review comments

There are high expectations for dialogue processing technology to realize smooth communication toward a human-robot symbiotic society.
Speech generation is realized using dialogue history and knowledge about the dialogue domain.
In this case, it is necessary to select knowledge appropriately from the knowledge sources that have been prepared in advance.
In this study, we investigated a method for selecting appropriate knowledge from the dialogue history under the condition that there is a mixture of standard information such as access and price that can be embedded directly in the utterance and non-standard information that cannot be embedded as itself, such as word of mouth and summary.
As a specific case study, we took a customer service dialogue in a travel agency and evaluated the effectiveness of the proposed method using 80 cases of questions collected in the Dialogue Robot Competition.