| 岡 佑依 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, | 渡辺 太郎, | 宮尾 知幸, | 須藤 克仁 |
|
title: Incorporating Noisy Length Constraints into Transformer with Length-aware Positional Encodings
abstract: Neural Machine Translation model often suffers from an under-translation due to its limited modeling of output sequence lengths. We propose a Transformer model using length constraints based on length-aware positional encoding. In training, we add random noise within a certain window size to the length constraints in the positional encoding. In inference step, we predict the output length using input sequence and a BERT-based length prediction model. Experimental results in ASPEC English to Japanese translation showed our method improved translation accuracy and controlled output sequence length. Additionally, we propose a method to incorporate length-aware positional encoding into Non-Autoregressive Translation model. language of the presentation: Japanese | ||||||
| 中野 佑哉 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, | 渡辺 太郎, | 宮尾 知幸, | 須藤 克仁 |
|
title: Solving Ambiguity on Question and Answering Tasks through Dialogue
abstract: Question answering is a task to find appropriate answers to given questions and present them, which is one of the basic and important tasks in various applications such as machine reading and interactive systems. Previous research on question answering systems has solved a variety of problems and achieved high accuracy on several benchmarks. However, there are some problems to be solved in the actual use of question answering systems. One of them is the ambiguity of user utterances to the question answering system. In this research, a new question answering task is designed to determine the unique meaning of a question by asking a question in response to an ambiguous question that does not have a unique answer. We then created a benchmark question-answer pairs' dataset for this problem by transforming an existing large dataset for the question answering task. We also conducted experiments to evaluate the accuracy of our dataset using existing models and discussed the problems that exist. language of the presentation: Japanese | ||||||
| 設樂 一碩 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, | 渡辺 太郎, | 田中 宏季 | |
|
title: Analysis of Verbal and Non-Verbal Behaviors during Cognitive Behavior Therapy thorough a Virtual Human Agent
abstract: Cognitive behavior therapy (CBT) is a dialogue based therapy that examines biased negative thoughts and seeks solutions to current problems. We aim to achieve low cost and highly effective automation with a virtual human agent. In this research, we created a fixed dialogue scenario based on the cognitive restructuring of CBT and implemented it in a virtual human agent. The agent was used for dialogue data collection with graduate and undergraduate students. Analysis of nonverbal behavior identified facial expressions that changed with the user's mood. Analysis of the varbal utterances also revealed the problem that some users were unable to properly state thoghts that caused user's negative mood. Based on the results, we considered the construction of dialogue modeling that coaching users to properly state the negative thoughts. language of the presentation: Japanese | ||||||
| 玉置 理沙 | M, 2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, | 金谷 重彦, | 藤本 まなと | |||
|
title:Construction of a blood glucose estimation model for the development of a blood glucose control system
abstract: In recent years, the prevalence of type 2 diabetes mellitus has increased in Japan due to the westernization of lifestyle-related diseases. However, there are few methods to support diabetic patients, and it is important to prevent the disease at an undiagnosed stage. For this purpose, it is necessary to control blood glucose levels, and a system to monitor blood glucose levels in real time is desirable. In this study, we analyzed data as the first step toward the realization of the system. Specifically, we collected blood glucose levels, dietary information, and sleep information from four subjects, constructed a blood glucose estimation model for each subject, and evaluated its accuracy using the RMSE, coefficient of determination, and MAE. language of the presentation: Japanese 発表題目: 血糖値コントロールシステム開発に向けた血糖値推定モデルの構築 発表概要: 型糖尿病の有病率が高くなっている。しかし、糖尿病患者を支援する手法はほとんど存在しなく、未病の段階で予防することが重要である。そのためには、血糖値をコントロールが必要であり、リアルタイムで血糖値を把握するシステムが望ましい。本研究では、システム実現に向けた第一歩としてデータ分析を行った。具体的には4名の被験者から血糖値と食事情報と睡眠情報を収集し、各被験者に対する血糖値推定モデルを構築し、RMSE、決定係数、MAEを用いてその精度を評価した。 | ||||||||
| 前川 哲志 | M, 2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, | 金谷 重彦, | 諏訪 博彦（特任准教授） | |||
|
title: Data collection and analysis for the Work Attitude PLR collection platform
タイトル: Work Attitude PLR収集基盤の構築に向けたデータ収集と分析 abstract: Japanese labor productivity is the lowest in the seven advanced countries(G7) despite the fact that many people are working such long hours with stress as to cause industrial accidents such as mental illness and suicide. Therefore, workstyle reforms that improve productivity while maintaining mental health are urgently needed. In the work style reform, health management based on "Work Attitude" such as stress, work engagement, and workaholism is essential, and it is advisable to measure daily Work Attitude in order to be able to detect the sign of the industrial accident quickly. However, the conventional Work Attitude measurement methods are based on a questionnaire once a year, and then they are not suitable for monitoring. Therefore, in this study, we develop Work Attitude PLR (Personal Life Record) collection platform that continuously measures and records Work Attitude that used to be measured only sporadically based on subjective questionnaires, using multimodal information. In this presentation, I describe the data collection experiments we conducted and their analysis for the construction of the proposed Work Attitude PLR collection platform. 概要: 我が国は，長時間労働やストレスにより多くの精神疾患患者や自殺者が発生しているにも関わらず，労働生産性は先進7カ国(G7)で最下位となっている．そのため，健康を維持したまま生産性を向上させるような働き方改革が急務とされている．働き方改革においては，労働者の好調・不調の程度を示す「Work Attitude」(ストレスやワーク・エンゲージメント，ワーカホリズムなど)に基づく健康管理は必須であり，日常的にWork Attitudeを計測することが求められている．一方で, 従来のWork Attitudeの評価手法は，1年に1回程度のアンケートによるものであり，モニタリングに適したものではない．そこで, 本研究では，主観的アンケートに基づいて散発的にのみ計測されていたWork Attitudeを，マルチモーダル情報を用いて持続的に計測・記録する「Work Attitude PLR（Personal Life Record）収集基盤」を構築する．本発表では,提案するWork Attitude PLR収集基盤の構築に向けて行ったデータ収集実験とその分析について述べる. language of the presentation: Japanese | ||||||||
| SALDAJENO DON PIETRO BAGADION | M, 2回目発表 | 数理情報学 | 池田 和司, | 金谷 重彦, | 吉本 潤一郎, | 久保 孝富（特任准教授）, | 福嶋 誠, | 日永田智絵 |
|
title: Elucidating the interactions between circadian transcription factors and E2F8 and NNMT through gene network inference
abstract: The circadian clock is a physiological clock present in most organisms, including humans, and is involved in the regulation of many essential bodily functions. Recent research suggests that the cancer-linked genes E2F8 and NNMT are controlled by the circadian clock. This suggests the possibility of targeting the transcription factors of the circadian clock with drugs for the purpose of cancer treatment. In order to perform such therapy, it is necessary to know the regulatory relationships between the circadian clock transcription factors, E2F8, and NNMT. However, the regulatory relationships between the circadian clock transcription factors, E2F8, and NNMT are not yet well-studied. In this research project, we attempt to elucidate these regulatory relationships through gene network inference using machine learning models. Two models were used. The first model uses ordinary differential equations and random forests to detect regulatory interactions. The second model uses Bayesian inference to predict the probabilities that regulatory interactions exist between certain genes. Using time-series gene expression data of ciracadian transciption factors, E2F8, and NNMT, gene regulatory networks were constructed with these two models. The results of each model, as well as plans for future work, will be discussed in this presentation. language of the presentation: English | ||||||||