Analysis of Conversational Listening Skills toward Agent-based Social Skills Training
Analysis of Conversational Listening Skills toward Agent-based Social Skills Training
复制标题
基于代理的社交技能训练的会话听力技能分析
DOI：
10.1007/s12193-019-00313-y
复制
发表时间：
2020
影响因子：
2.9
通讯作者：
Satoshi Nakamura
中科院分区：
文献类型：
--
作者：
Hiroki Tanaka;Hidemi Iwasaka;Hideki Negoro;Satoshi Nakamura
倾听技巧对于人类交流至关重要。社会技能训练（SST），由人类教练执行，是一种行之有效的方法，以获得适当的技能，在社会交往。以前的工作通过开发一个对话系统，通过与计算机代理的交互来教授口语技能，从而实现了社交技能培训过程的自动化。尽管之前的模拟社交技能训练的工作考虑了说话技能，但SST框架还包含了其他技能，如倾听、提问和表达不适。在本文中，我们扩展了我们的自动化的社交技能培训，考虑用户在与计算机代理的对话中的倾听技巧。我们准备了两个场景：听力1和听力2，分别假设闲聊和工作培训。一位女特工向参与者讲述了最近的一个故事，以及如何打电话，参与者听着。我们记录了27名日本研究生与代理人互动的数据。两位外部专家评估了参与者的听力技能。我们手动提取了可能与参与者的眼睛注视和行为线索相关的特征，并证实了一个简单的线性回归与选定的特征正确地预测了听力技能，在这两种情况下的相关系数都在0.50以上。话语中的重复和反向通道的数量有助于预测，因为我们发现，仅使用这两个特征就可以预测听力技能，相关系数高于0.43。由于这两个功能对用户来说更容易理解，我们计划将它们集成到自动化社交技能培训的框架中。
Listening skills are critical for human communication. Social skills training (SST), performed by human trainers, is a well-established method for obtaining appropriate skills in social interaction. Previous work automated the process of social skills training by developing a dialogue system that teaches speaking skills through interaction with a computer agent. Even though previous work that simulated social skills training considered speaking skills, the SST framework incorporates other skills, such as listening, asking questions, and expressing discomfort. In this paper, we extend our automated social skills training by considering user listening skills during conversations with computer agents. We prepared two scenarios: Listening 1 and Listening 2, which respectively assume small talk and job training. A female agent spoke to the participants about a recent story and how to make a telephone call, and the participants listened. We recorded the data of 27 Japanese graduate students who interacted with the agent. Two expert external raters assessed the participants’ listening skills. We manually extracted features that might be related to the eye fixation and behavioral cues of the participants and confirmed that a simple linear regression with selected features correctly predicted listening skills with a correlation coefficient above 0.50 in both scenarios. The number of noddings and backchannels within the utterances contributes to the predictions because we found that just using these two features predicted listening skills with a correlation coefficient above 0.43. Since these two features are easier to understand for users, we plan to integrate them into the framework of automated social skills training.