2021 Fiscal Year Annual Research Report
Learning conversation agents for long-vision response generation in multi-round communications
Project/Area Number |
19K20345
|
Research Institution | The University of Tokushima |
Principal Investigator |
康 シン 徳島大学, 大学院社会産業理工学研究部(理工学域), 助教 (80777350)
|
Project Period (FY) |
2019-04-01 – 2022-03-31
|
Keywords | long-vision conversation / dialogue topic tagging / neuro-symbolic AI / DialogQuality Evaluation |
Outline of Annual Research Achievements |
We refined the automatic tagging of dialogue- and utterance-level topics to Japanese Twitter conversations and proposed a neuro-symbolic model for the automatic dialogue quality evaluation. We improved the strategic conversation learning procedure by training the utterance generator with a precise and meaningful topic conversation corpus and training a dialogue quality discriminator with a neuro-symbolic model. Specifically, we built the topic tagging system by encoding dialogues, utterances, and words in the same semantic space, took cluster centers as topics, and created a topic-transition diagram. We proposed 3 criteria for evaluating the match of topics and dialogues. Results suggested that the clusters indicated meaningful topics and precise match of topics and dialogues were obtained by restricting the lower bound of utterance length in 10 to 20. We proposed a neuro-symbolic method for automatic dialogue quality evaluation, in which speaker identity, position of utterance, and dialogue topics were provided as the symbolic knowledge to facilitate a transformer-based model. The method achieved the best dialogue quality evaluation result in NTCIR-16 DialEval-2 Task. We developed the strategic conversation learning procedure based on the refined topic-conversation corpus and the automatic dialogue quality evaluator. Our preliminary study suggested that matching of topic and utterance was essential to improve the succeeding rate in generating strategic conversations and that the neuro-symbolic method was helpful to detect logical flaws in the generated conversations.
|