| 長村 佳歩 | 1651027: M, 2回目発表 | 知能コミュニケーション 中村 哲, 松本 裕治, Sakriani Sakti, 須藤 克仁 |
|
title:
Neural Machine Translation Considering ASR Errors
abstract: Spoken language translation system consists by at least two modules: an automatic speech recognition (ASR) system and a machine translation (MT) system. It is well known that there is a high correlation between ASR errors and translation quality. If there are some errors in the result of ASR, the accuracy of the final translation result decreases significantly. Up to now, there have been many efforts to construct tightly-coupled ASR and phrase-based MT systems that are jointly trained and optimized. But it is still only a few researches that focus on currently trending neural machine translation (NMT). This research focuses on maintaining high accuracy of NMT. Specifically, we aim to realize robust speech translation to ASR errors by learning ASR error tendency in NMT and estimating optimum output. In this presentation, we will describe our efforts and future research directions. language of the presentation: Japanese 発表題目: 音声認識誤りを考慮したニューラル機械翻訳 発表概要: 音声翻訳システムは音声認識と機械翻訳，テキスト音声合成の３つのモジュールで構成されている．現行の音声翻訳では，音声認識器の結果に誤りがあると機械翻訳器の翻訳結果が連鎖的に誤り，最終的な翻訳結果の精度が大幅に低下する．そこで，本研究では高い翻訳精度を維持するために，機械翻訳器に音声認識誤り傾向を学習させ，最適な出力の推定を行うことで，音声認識誤りに頑健な音声翻訳の実現を目指す．また，本発表では今までの取り組みと今後の研究の方針について述べる． | ||
| 細見 直希 | 1651094: M, 2回目発表 | 知能コミュニケーション 中村 哲, 松本 裕治, Sakriani Sakti, 吉野 幸一郎 |
|
title: Deception detection and analysis from dialogue by using ensemble learning
abstract: Deception often occurs in human-human communication. However, several research works on lie detection help us learn that it is not easy for humans to detect it. This problem has led researchers to look at automated ways of assessing the truth value by using a computer. Recently, machine learning methods have been extensively studied, and especially in lexical features, the classifier which uses distributed representations called "fastText" achieved high performance in sentiment analysis and tag prediction tasks. In this work, we utilize "fastText" approach for a deception classifier. Furthermore, we constructed ensemble classifier based on text and speech features, as well as analyzed deception detection accuracy in comparison with human performance. The results show that it is possible to construct a deception classifier which is significantly higher precision than human. language of the presentation: Japanese | ||
| 森 巧磨 | 1651110: M, 2回目発表 | 知能コミュニケーション 中村 哲, 松本 裕治, Sakriani Sakti |
|
title: Model compression of end-to-end automatic speech recognition by tensor train and sequence knowledge distillation
abstract: Automatic speech recognition (ASR) system converts human speech into words. It is a major component of many applications such as speech summarization, automatic call center, and speech-to-speech translation. Recent technologies used in ASR are based on an end-to-end framework using deep learning. However, most of the neural network structures have millions of parameters and require many computational resources for training and predicting new data. This research proposes an alternative neural network model to reduce the number of parameters significantly by representing the weight parameters based on Tensor Train format. Furthermore, we also use knowledge-distillation method which trained the compressed student network from the original teacher network to address the performance gap. The results show that we could minimize the number of parameters while preserving the ASR performance. language of the presentation: Japanese | ||
| 叶 高朋 | 1661005: D, 中間発表 | 知能コミュニケーション 中村 哲, 松本 裕治, Sakriani Sakti |
|
title: Structured-based Curriculum Learning for\\End-to-end English-Japanese Speech Translation
abstract:Sequence-to-sequence attentional-based neural network architectures have been shown to provide a powerful model for machine translation and speech recognition. Recently, several works have attempted to extend the models for end-to-end speech translation task. However, the usefulness of these models were only investigated on language pairs with similar syntax and word order (e.g., English-French or English-Spanish). In this work, we focus on end-to-end speech translation tasks on syntactically distant language pairs (e.g., English-Japanese) that require distant word reordering. To guide the encoder-decoder attentional model to learn this difficult problem, we propose a structured-based curriculum learning strategy. Unlike conventional curriculum learning that gradually emphasizes difficult data examples, we formalize learning strategies from easier network structures to more difficult network structures. Here, we start the training with end-to-end encoder-decoder for speech recognition or text-based machine translation task then gradually move to end-to-end speech translation task. The experiment results show that the proposed approach could provide significant improvements in comparison with the one without curriculum learning. language of the presentation: Japanese | ||
| 梶原 武紘 | 1651032: M, 2回目発表 | 光メディアインタフェース 向川 康博, 佐藤 嘉伸, 舩冨 卓哉, 久保 尋之, 田中 賢一郎 |
|
title: Non-Rigid Registration of Serial Section Images by Blending Rigid Transformations
abstract: We describe feature-based non-rigid registration of histological serial section images. Our method represents non-rigid deformation by blending the rigid transformations estimated in the local region around a control point. This approach can efficiently represent non-rigid deformation with a smaller number of control points than conventional methods that interpolate displacement, such as free-form deformation (FFD). A feature-based approach is adopted to extract the control points and robustly estimate the local rigid transformation at each control point. By blending the rigid transformations, the transformation at each pixel is computed as a transformation field. The experimental results demonstrate that the proposed method is effective for achieving non-rigid registration efficiently and robustly for histological serial section images. We also introduce the way to solve the problem in changing the area of tissue by non-rigid registration. language of the presentation: Japanese | ||
| 西川 雅清 | 1651082: M, 2回目発表 | 光メディアインタフェース 向川 康博, 佐藤 嘉伸, 舩冨 卓哉, 久保 尋之, 田中 賢一郎 |
|
title: Visualizing Displacement of Rays caused by Refraction through Cells in Contact Imaging
abstract: We are developing novel imaging system called lens-less microscope which is downsized microscope by using contact imaging which captures cells placed on the image sensor directly and they are illuminated by a point light source above the sensor. In this research, we propose measuring method of the displacement of rays caused by refraction through cells to facilitate observation of the cell structure in contact imaging by showing this novel information which depends on the angle of the cell surface. The displacement of rays is computed by variation of arrival position of the ray on the sensor using input images captured while shifting the stripe mask inserted between point light source and cells. We confirmed that we can compute the displacement of rays by experiment designed in imitation of the contact imaging and we visualize displacement value using pseudo color. In addition, we made experiment using sea urchin fertilized egg cells images captured by the lens-less microscope to verify the effectiveness of our method. language of the presentation: Japanese 発表題目: コンタクトイメージングにおける細胞による屈折変位量の可視化 発表概要: 我々は，イメージセンサ上に直接細胞を置き，上部の点光源により細胞を照らして撮影するコンタクトイメージングを用いたレンズレス顕微鏡を開発している．本研究では，コンタクトイメージングにおける細胞の表面形状や内部構造の観察を容易にするため，細胞での光線の屈折による変位量を撮像する手法を提案する．屈折変位量の主要因は細胞表面との入射角度に応じて生じる屈折現象であり，観察時にこの新たな情報を提示することで細胞の立体構造の把握の手助けになると考えられる．屈折変位量は，点光源と細胞の間に挿入したストライプマスクをシフトさせながら撮影した入力画像群を用いて，イメージセンサ上での光線が到達する位置の変動から算出する．コンタクトイメージングを模した環境での実験により屈折変位量が算出できることを確認し，屈折変位量を疑似カラーで可視化した．さらに，実際のレンズレス顕微鏡で撮影したウニの受精卵細胞に対して実験を行い，提案手法の有効性を検証した． | ||
| 木戸 勇太 | 1651043: 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, 荒川 豊, 水本 旭洋 |
|
title: Seasoning quantity adjustment support system considering user's preference
abstract: In recent years, online recipe sites such as Cook Pad 揃 Rakuten recipe have increased, and people who cook while observing the recipe are increasing year by year. However, the online recipe does not necessarily correspond to seasoning for each of many users. In addition, it is known that 50% or less of persons measuring and injecting seasoning properly at seasoning, many people who have not cooked according to the taste preference of the user, that is, the taste. Therefore, in my research, I analyze user 's taste, adjust the amount of recipe according to taste, and propose a system that intuitively understand the appropriate amount in real time during input. In this presentation, we will describe the progress so far and future issues. language of the presentation:Japanese | ||
| 雨森 千周 | 1651003: 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, 荒川 豊, 諏訪 博彦, 水本 旭洋 |
|
title: Simplified HRQOL Measurement Method by using Smart Device
abstract: Health-related quality of life (HRQOL) is a useful indicator that rates a person's activities in various physical, mental and social domains. Continuously measuring HRQOL can help detect the early signs of declines in these activities and lead to steps to prevent such declines. However, it is difficult to continuously measure HRQOL by conventional methods, since its measurement requires each user to answer burdensome questionnaires. In this paper, we propose a simplified HRQOL measurement method for a continuous HRQOL measurement which can reduce the burden of questionnaires. In our method, sensor data from smart devices and the questionnaire scores of HRQOL are collected and used to construct a machine-learning model that estimates the score for each HRQOL questionnaire item. Our experiment result showed our methodソs potential and found effective features for some questions. language of the presentation: Japanese | ||
| 梅木 寿人 | 1651020: 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, 笠原 正治, 荒川 豊, 藤本 まなと, 水本 旭洋 |
|
title:Determination of evacuation location method in congested area at time of disaster
abstract: In Japan, evacuation training / induction is one of the serious disaster prevention / reduction measures.At present, evacuation guidance at the time of a disaster is to guide you from the present location to the nearest evacuation center. In the situation where many people evacuate all together, the present evacuation guidance does not consider delay of evacuation caused by congestion occurring on the route to evacuation centers, exceeding the shelters of maximum storage capacity. In this paper, we propose an evacuation guidance method that minimizes the sum of the time required for evacuation by all people, considering the shelters of maximum storage capacity in congested areas such as tourist spots and large cities, and the degree of congestion on meridians . Assuming that the degree of congestion in each area at the time of a disaster is known, the proposed evacuation guidance system presents evacuation places and evacuation route to people. In order to consider the shelters of maximum storage capacity, solve multiple snap problems, determine the initial evacuation center, simulate the congestion on the evaluation route and estimate the evacuation time considering congestion by simulation. A semi-optimal solution is obtained by repeating a simulation such as re-determining the evacuation destination so as to reduce the sum of the evacuation time of all people. We report on the results of the performance evaluation scenario of the proposed method on the simulator and the proposed method. language of the presentation: Japanese | ||