コロキアムB発表

日時: 9月26日(木)3限(13:30~15:00)


会場: L1

司会: Doudou Fall
堀口 博 M, 2回目発表 自然言語処理学 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之
title: Entity linking using knowledge graph structure information
abstract: In the entity linking that has been performed under the conventional Wikipedia domain, it is possible to acquire a mention-entity pair from a large number of hyperlinks, and it is possible to acquire an entity embedding that takes text information into account. However, it is not obvious that there are such a large number of mention-entity pairs. In this study, we propose to use entity embedding encoded from the structural information of knowledge base for linking. As a result, it was suggested that the knowledge graph possesses information to assist the text and the possibility of improvement of candidate generation using entity embedding.
language of the presentation: Japanese
 
山元 勇輝 M, 2回目発表 自然言語処理学 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之
title: Extending AMR with Generative Lexicon
abstract: In order for a computer to capture the semantics of natural language, we need a semantic representation. Abstract Meaning Representation (AMR) is a symbolic semantic representation which encodes meaning of a sentence in a directed acyclic graph: nodes represent abstract concepts and edges represent relations between them. In this research, we make an extension to the structure of AMR based on the idea of Generative Lexicon Theory. We then propose a semantic parser which converts a sentence into the extended format.
language of the presentation: Japanese
 
ZHAO XIN M, 2回目発表 自然言語処理学 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之
title: *** Unsupervised entity alignment via optimal transport ***
abstract: *** Different knowledge graphs usually contain the same or similar entities. Entity alignment information benefits knowledge graph construction and connection. Most of the recent solutions for entity alignment are based on knowledge graph embedding, which provides methods to align entities efficiently. However, they still need large size of seed dictionary. We intend to propose a new approach based on optimal transport problem to reduce reliance on supervised data. ***
language of the presentation: *** Japanese ***
 
柳田 智也 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: Toward Syntheized Speech in Machine incremental speech-to-speech interpreters with Neural Text-to-Speech Framework
abstract: Real-time machine speech interpreters aim to mimic human interpreters that able to produce high-quality speech translations on the fly. It requires all system components, including speech recognition, machine translation, and text-to-speech (TTS), to perform incrementally before the speaker has spoken an entire sentence. Existing studies of incremental TTS (iTTS) have mainly been conducted on a model based on hidden Markov model (HMM). Recently, end-to-end TTS based on a neural net has synthesized more natural speech than HMM-based systems. In this presentation, we take an initial step to construct iTTS based on end-to-end neural framework (Neural iTTS) and investigate the effects of various incremental units on the quality of end-to-end neural speech synthesis.
language of the presentation: Japanese
 

会場: L2

司会: Alexander Plopski
清川 拓哉 D, 中間発表 ロボティクス 小笠原 司, 加藤 博一, 高松 淳, 丁 明, Gustavo Garcia
title: Agile Assembly Robot
abstract: Agile manufacturing system using autonomous robots is expected to largely contribute to the high-mix low-volume production, a crucial paradigm shift in the industry field. The research purpose is to construct a framework to improve the time-consuming and laborious process in the case of using human-annotated datasets and human demonstrations for robot teaching. In this presentation, I will report the experimental results mainly about the proposed efficient method of collecting object detection datasets and object observation-based assembly planning.
language of the presentation: Japanese
 
永島 幹久 M, 2回目発表 ロボティクス 小笠原 司, 加藤 博一, 高松 淳, 丁 明
title: Prediction of the plantar force for the gait assistance using a lower-limb exskeleton
abstract: Based on the predicted gait status, the exoskeleton could be controlled to move the legs with less time delay. In this research, a method is proposed to predict future plantar force during the gait using IMU and plantar sensors. Based on the predicted plantar force, the gait status can be predict to control the exoskeleton without any delay. Deep neural networks (DNN) and long short-term memory (LSTM) are tested to learn the non-linear relationship between the measured sensor data and the future plantar force data. Using the trained network, the plantar force can be predicted not only during walking but also at the start and end of walking. In this presentation, I will introduce the method and show the result and the performance for different prediction time.
language of the presentation: Japanese
 
三上 徹朗 M, 2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, 久保 尋之
title: Measurement of subsurface light transport based on synchronization delay of projector camera system
abstract: Capturing the light transport by the projector-camera system, it is possible to acquire the light which passes under the surface.Light transport according to the depth under the surface can be obtained by changing the distance between the light illumination position and the observation position.In this study, I capture the light transport according to the depth under the surface by changing the distance between the light illumination position and the observation position. To show that the light transport under the surface differs depending on the wavelength of light , I also constructed a device that can shoot by replacing the light source of the laser projector from visible light to near infrared light. In this presentation, I show the visualization subsurface images obtained using the constructed device and future issues.
language of the presentation: *** Japanese ***
 
山田 憲 M, 2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, 久保 尋之田中 賢一郎
title: Shape measurement of concave metal by projection code using temporal frequency
abstract: Metals are used in various places and products in our daily life. For this reason, it can be said that there is a great demand for measuring the three-dimensional shape. However, in active stereo measurement, the illumination patterns overlap due to inter reflection, making it difficult to capture the shape of metallic object. In this presentation, I will propose a method to reduce the influence of inter reflection by separating the superimposed pattern in Fourier domain by using temporal frequency in the illumination pattern. Through experiments, we show that the 3D shape of the concave metal surface can be measured even if there is a region where the inter reflection is dominant. Finally, I will explain the future work of this research.
language of the presentation: Japanese
発表題目: 時間的周波数を利用した投影コードによる凹面金属の形状計測
発表概要: 金属は今日の我々の日常生活において様々な場所や製品に用いられており,その三次元形状を測ることに大きな需要があると言える.しかし,金属は材質の特徴から凹面に強い相互反射が発生するため,アクティブ照明法を用いた形状計測において照明パターンが重畳し,その形状を捉えることを困難にしている.本発表では,その課題を解決すべく今まで取り組んできた研究成果として,照明パターンに時間周波数を使用することで重畳したパターンをフーリエ空間で分離し相互反射の影響を低減させる方法を提案する.実験を通し,相互反射が支配的な部位があっても,金属凹面の3次元形状を測定できることを示す.最後に本研究の今後の展望に関して述べていく.