| 上野 友梨 | M, 1回目発表 | 光メディアインタフェース | 向川 康博, | 加藤 博一, | 舩冨 卓哉, | 田中 賢一郎 |
|
title:
The front and back ink painting image on one piece of paper modeling with two transmittances and scattering of light for Literature research
abstract: This research will make the front and back ink painting image on one piece of paper modeling with transmission to make the Literature documents more readable. The last goal is to separate the front and back ink painting image, and revealed the light phenomenon within paper in this step. In this research, we focused on the paper structure, which is constructed from two filtering layers and the blur layer. As considering this structure, we created two pattern modelings with two type light transmittances in the front and back aspects and the light scattering in the middle layer of paper. The light will be decreased through the filtering layers with ink painting area, and broken up within the blur layer. Introducing this idea, we simulated and compared with the simulation result and the photo in the real situation in order to confirm the quality of our modeling. This comparisons suggest the simulation works well to some extent, however there is left a little room to improve the simulation and modeling more for the variation of paper. To obtain sufficient image data in various cases can make our proposed modeling suitable to separate the front and back image on one piece of paper in the future. language of the presentation: Japanese | ||||||
| CHOTCHAICHARIN SETTHAWUT | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清, | 加藤 博一, | 酒田 信親, | 磯山 直也 |
|
title: AR Real-time Earthquake Simulation
abstract: There are over 100 reports of earthquakes in Japanese history. These earthquake disasters can cause severe loss of life and economy. Various media exist to simulate and educate people about earthquakes, but they are a number of limitations. First, most media uses the setup screen, which has a significant impact on the user's mind but only for the first time. Second, they limit the user's mobility and reduce the realism of the simulation. To overcome these limitations, we propose an AR earthquake simulation system using a video see-through head mounted display that can generate a virtual earthquake in any indoor environment. We plan to acquire furniture objects in the real environment with the help of an object database. We then plan to diminish the real furniture and overlay the corresponding virtual ones on the video background. Finally, we plan to perform earthquake simulation by using real acceleration data of a past earthquake event such as the 2011 Tohoku-Pacific Ocean Earthquake. By adding force data to the virtual objects, a user can experience a virtual earthquake in the real environment. We also plan to use galvanic vestibular sensations for multi-modal earthquake simulation. language of the presentation: English | ||||||
| YANG ZHENGCHANG | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清, | 加藤 博一, | 酒田 信親, | 磯山 直也 |
|
Title: Palm-Based See-Through Interface for Selection in the AR Environment Abstract: Object selection has been identified as a fundamental task in 3D user interface. Selection techniques, such as Raycasting, head-pointing and gaze, are more and more reliable. However, AR and VR still encounter difficulties in small far object selection and object occlusion. Current selection techniques keep visual feedback but lose haptic feedback which can increase accuracy of selection. To address these problem, we propose a new on-body see-through selection interface. We transfer non-dominant hand as a see-through touch device. The interface employs corresponding hand movement to interact with virtual objects, dealing with object occlusion, while provides haptic feedback. We'll implement and conduct user study in the future. Language of the presentation: English | ||||||
| 柳井 丈司 | M, 1回目発表 | 知能コミュニケーション | 中村 哲☆3, | 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: One of the purposes of open-domain dialog systems that conduct chit-chat and question answering is to offer varieties of knowledge ranging from daily chores to business issues through system responses. In this research, we propose a method for ranking response candidates using a BERT-based topic word model to select the response candidates that are relevant to user interests, which we assume that the system has identified through simple question-answering with the user. Furthermore, to enable the system to provide a wide variety of responses that match user interests, we propose a method to iteratively collect response candidates from a large web archive. In this research, we use WEKDA (Mizuno et al., 2018), which is a dialog system developed by NICT, as a base system. WEKDA aims at providing future opportunities or risks to the user and gives the responses that represent the causal consequences of users' inputs. For example, in response to the user's input such as "I'm concerned about global warming?", WEKDA outputs a response such as "Global warming seems to cause worldwide climate change," using the causal relation between "global warming worsens" and "worldwide climate change occurs." A problem is that a collected set of response candidates do not always contain a candidate relevant to the user interests by following such a causal relation only once. In the proposed method, we follow such a causal relation iteratively and select the system response from the collected larger set of response candidates. We expect that this makes it more likely to provide system responses that match the user interests. Finally, we combine two proposed methods. We validate the effectiveness of each proposed method through the subjective evaluation, in which we ask multiple annotators whether the system responses generated by our proposed methods have higher quality than those produced by the baseline method, i.e., the current WEKDA's response generation method. language of the presentation: Japanese 発表題目: オープンドメイン対話システムにおけるユーザの興味を用いた応答候補選択 発表概要: 大量のweb文書を基に雑談や質問応答を行うオープンドメイン対話システムの目的の一つに、ユーザの興味を反映したシステム応答により、ユーザに日常生活から仕事まで幅広い内容に関する新たな知識を提供することがある。本研究では、事前に質問等によって取得したユーザの興味の対象を示す単語等を基にして、BERTベースのトピックワードモデルを使用してシステムの応答候補をランキングし、よりユーザの興味に関連の深い応答を選択する手法を提案する。さらに、システムが応答候補のバリエーションを増やして、よりユーザの興味に関連の深い情報を提供可能とするために、ユーザ発話を受けて、Webアーカイブ等の大規模テキストデータベースから一度、応答候補を収集したのち、因果関係の連鎖に基づいて追加の応答候補を収集する方法を提案する。本研究では、ベースとなる対話システムとしてNICTが開発したWEKDA (Mizuno et al. 2018)を利用する。このシステムは、将来におけるチャンスやリスクをユーザに提供することを目的とし、様々な話題に関して、ユーザ入力に対して因果関係が成り立つ可能性があるとシステムが判断した出来事を記述する応答候補を応答として出力する。例えば、ユーザ発話「地球温暖化が心配だ」に対して、「A: 地球温暖化が進む」と「B: 気候が変動する」のような「AによりBが生じる」関係を成立させる因果関係に基づいて、「温暖化は地球規模の気候変動をもたらすらしい。」といった応答を出力する。ただし、因果関係を一回探索することで得られる発話候補集合にユーザの興味に合致する発話が必ずしも含まれているとは限らない。そこで本手法では、因果関係を複数回探索し、その結果得られるより大量の応答候補集合からシステム応答を選択する。これにより、ユーザの興味に合致する発話を提供できる可能性が高くなると考えられる。最終的に、本研究で提案する二つの手法を組み合わせ、ユーザの興味との関連度を推定するBERTベースのトピックワードモデルを大量に収集した応答候補に適用することで、ユーザの興味に関連の深い可能性が高い応答ができるようになると考えられる。それぞれの提案手法について、提案手法を用いた場合のシステム応答と、提案手法を用いない場合のシステム応答とでどちらが応答としてふさわしいか、アノテータによる主観評価を行い、提案手法の有効性を検証する。 | ||||||
| LENKA PABITRA | M, 1回目発表 | 知能コミュニケーション | 中村 哲☆3, | 荒牧 英治, | 鳥澤 健太郎（客員教授）, | 飯田 龍（客員准教授） |
| Title: Paraphrasing Questions for Question Answering Systems Abstract: Questions given to a Question Answering (QA) system are an important information source for finding answers. However, a QA system often fails to find answers if the input questions are too vague for the QA system to collect evidence from passages. In such a case, a QA system needs to reformulate information of the input questions based on its understanding and use the reformulated information for collecting evidence for answers. We assume that giving input questions along with their paraphrased questions to QA systems could improve their performance. Paraphrasing consists in generating sentences or questions using different keywords without changing its original meaning. For example, a QA system may fail to provide a correct answer to the question “Why do cats worry about cucumbers?” possibly due to the limiting reach of the keyword “worry” in collecting evidence from passages. To tackle the problem, we obtain additional keywords like “fear”, “scared”, and “feel threatened” from multiple paraphrased questions like “Why do cats fear cucumbers?”, “Why are cats scared of cucumbers?”, and “Why do cats feel threatened by cucumbers?”. The additional keywords can guide QA systems in collecting better evidence from passages and extracting better answer spans. We work on this task in the following two steps: 1) generating multiple paraphrases for a given question using a paraphrase generator and 2) extracting answer for the given question using a QA model. Language of the presentation: English | ||||||