| 河野 誠也 | D, 中間発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 吉野 幸一郎 |
|
title: Neural Conversation Model Controllable by Given Dialogue Act Based on Adversarial Learning and Label-aware Objective
abstract: Building a controllable neural conversation model (NCM) is an important task. NCMs which learn a direct mapping between a dialogue history and a response utterance, are widely researched as a flexible approach to building non-task oriented dialogue systems. However, it is difficult to control their responses on the basis of actual constraints such as dialogue act classes. In this work, we focus on controlling the responses of NCMs by using dialogue act labels of responses as conditions. We introduce an adversarial learning framework for the task of generating conditional responses with a new objective to a discriminator, which explicitly distinguishes sentences by using labels. This approach strongly encourages the generation of label-conditioned sentences. We compared the proposed method with some existing methods for generating conditional responses. The experimental results show that our proposed method has higher controllability for dialogue acts even though it has higher or comparable naturalness to existing methods. In this presentation, we also discuss the future direction to incorporate the actual dialogue phenomena into the proposed model. language of the presentation: Japanese | |||
| 村瀬 行俊 | D, 中間発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 吉野 幸一郎 |
|
title:
Engaging Users in Dialog System with Commonsense Knowledge
abstract: The goal of this research is improving user engagement in dialog systems with a commonsense knowledge base (CSK-Base). Previous dialog system research attempts to build inference models for specific domains or embed commonsense inference in sequence-to-sequence models. However, a commonsense inference task on the dialog to measure plausible inference in a dialog system has not yet to be investigated. We address this problem by proposing to create a dataset where commonsense inference on CSK-Base is annotated on the dialog. On the dataset, we will develop more efficient a commonsense inference module for the dialog system. In this presentation, we focus on the annotation schema test with a small number of dialog sessions. We provide the analysis of the inter-coder agreement, which indicates how people agree with the commonsense inference on CSK-Base. We also investigate the decision boundaries of plausible commonsense inferences. We consider the means and medians of annotators' scores as the boundaries. We discuss the properties of them for formalizing a commonsense inference task. language of the presentation: Japanese | |||