| 柳井 丈司 | M, 2回目発表 | 知能コミュニケーション | 中村 哲☆, | 荒牧 英治, | 鳥澤 健太郎（客員教授）, | 飯田 龍（客員准教授） |
|
title: Response Candidate Selection Using User Interests in Open-Domain Dialog Systems
abstract: Response selection is a crucial part of retrieval-based dialogue systems. Incorporating user's interests into response selection is expected to increase user engagement. Prior works consider the response selection as a ranking problem and used various information as the user's interests. In this study, we propose a method of response selection based on the past utterances of the user, considering the past utterances as the user's interests. We calculate the semantic similarity between user's interests and response candidates by using the output of BERT-based model. This model estimates the topic of input sentences. Then we select the best response from response candidates based on the similarity score. We validate the effectiveness of our proposed method through subjective evaluation, where multiple annotators judge if the responses selected by our proposed method are relevant to the user's interests. Since it is expected that individual annotation decisions may vary from person to person, multiple annotators join the evaluation. language of the presentation: Japanese | ||||||
| LENKA PABITRA | M, 2回目発表 | 知能コミュニケーション | 中村 哲☆, | 荒牧 英治, | 鳥澤 健太郎（客員教授）, | 飯田 龍（客員准教授） |
|
title: Automatic Assessment of Argument Convincingness in Online Debates
abstract: Argument convincingness assessment is a task to assess the degree of convincingness of an argument for a given debatable topic. In order to have a fruitful debate with a human, a system should come up with effective and convincing arguments even for the most delicate topics. Previous work pointed out that arguments which are long, contain more details, more reasons, and facts tend to be more convincing. However, they did not consider the possibility of causalities to judge the convincingness of arguments. A causality or causal relation is a relation between two events such that one is the effect of the other. For example, for the given topic "We should adopt open-source software", if any of its argument contains the effect of adopting open-source software then it tends to be more convincing. Our manual analysis on a subset of arguments also revealed that causality is an important indicator to argument convincingness. Inspired from this analysis and the prevalent use of causalities in our daily conversations and arguments, we assume that causalities will be useful to find more convincing arguments. Therefore, to tackle this task we propose the following two methods 1) detect causality between a topic and a pair of topic-related arguments using a causality detector and 2) use the causalities and assess argument convincingness using a convincingness classifier. Our preliminary experiments show that adding causalities to baseline did not outperform the baseline. We are now conducting error analysis and trying to develop a more effective method using the causalities. language of the presentation: English | ||||||
| 吉川 莉央 | M, 1回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 荒牧 英治, | 諏訪 博彦, | 松田裕貴 |
|
title: An investigation of the environmental factors' effects on responses reliability in participatory sensing
abstract: In participatory sensing, users do not always give accurate responses (careless responses), and the quality of the data might not be guaranteed. Related studies have pointed out that careless responses are caused by the attitude of the user who does not pay appropriate cognitive costs for the given task. In this study, we hypothesized that not only the user's attitude toward the task but also stressful environmental factors surrounding the user (e.g., time constraints, environmental noise) may cause careless responses. In this presentation, I will report the results of investigating the status of careless responses (correct response rate and response behavior) to a participatory sensing task under both stressful and normal conditions. language of the presentation: Japanese 発表題目: 環境要因が参加型センシングの回答信頼性に与える影響調査 発表概要: 参加型センシングにおいて，ユーザは必ずしも正確な回答をするとは限らず（不良回答），データの質の保証ができないという問題がある． これまでの研究では，不良回答は与えられたタスクに対して適切な認知コストを払わない協力者の態度によって引き起こされると指摘されている． 本研究では，タスクに対するユーザの態度だけでなく，ユーザを取り巻くストレスフルな環境要因（例えば，時間制約や騒音など）も不良回答を引き起こすのではないかと考えた． 本発表では，ストレスの多い状況下と通常時において参加型センシングタスクの不良回答状況（正答率や回答状況）を調査した結果について報告する． | ||||||
| ZHANG YUJIA | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清, | 向川 康博, | 酒田 信親, | 磯山 直也 |
| title: Effect of real shadow in images on virtual objects in augmented reality abstract: When inserting a virtual object into a picture or video, the virtual object will always be affected by the lighting environment in the image. In addition, the virtual object will also be affected by the shadows that already exist in the image. Our research is about the effect of shadow in the image on virtual object. Especially the change of its surface when the virtual object passes the edge of the shadow. language of the presentation: Japanese | ||||||
| HU XIAODAN | M, 2回目発表 | サイバネティクス・リアリティ工学 | 清川 清, | 向川 康博, | 酒田 信親, | 磯山 直也 |
|
title: Computational Sunglasses for Autism Spectrum Disorders
abstract: We design computational sunglasses as an assistant device for individuals with autism spectrum disorders(ASDs) to alleviate the hypersensitive visual perception causing by abnormal neuronal development. Such abnormal development leads to atypical pupil mechanisms such as larger static baseline pupil size and longer pupillary light reflex, all of these reveal that individuals with ASD encounter the symptom as photophobia. Basing on the atypical pupil mechanisms, we use a camera to simulate their large pupil size and propose a system with a scene camera for ambient illuminance detection, and a single spatial light modulator (SLM) which is capable of per-pixel transmittance adjustment by applying a control signal with the response of sensor irradiance of the scene camera. However, the occlusion, which possesses low transmittance to block the incoming light ray, becomes blurred since the focal plane is not on the SLM and blocks the light stimulation insufficiently. We do not intend to add extra optics to coincide the focal plane and the SLM plane since that will make the sunglasses bulky while the individuals with ASD are hypersensitive in perception. Therefore, we optimize the occlusion mask from an image processing perspective. We define the appropriate occlusion for our system, simulate the imaging process with an out-of-focus point spread function, and finally obtain the optimal parameters for the occlusion mask. The optimal approach demonstrates better scene contrast evaluated by the camera system. language of the presentation: English | ||||||