日時: 9月24日(月)5限 (16:50-18:20)

会場: L1

司会: Sakriani Sakti
程飛 1151131: M, 2回目発表 松本裕治, 中村哲, 新保仁, Kevin Duh
title: Parsing morphological structure of Chinese Synthetic words with Character-based Dependency Model
abstract: We propose a character-based dependency model with our custom morphological dependency relation label set to parsing Chinese synthetic words into a dependency tree. By these morphological relation labels, we can transform almost any kind of dependency tree to structure tree for Chinese synthetic word.
language of the presentation: English
江崎大嗣 1151134: M, 2回目発表 松本裕治, 中村哲, 新保仁, Kevin Duh
title: Multiple Emotion classification using pairs of emotions
abstract: Estimating emotion in sentence is required because it is usefule for dialogue system and customer analysis. In the conventional method, they use classifier such as SVM in emotion classification, and simply classify emotion as multiple values. In this presentation, I describe the new method that consider a pair of emotons as a axis in emotion classification.
language of the presentation: Japanese
発表題目: 感情対を用いた多クラス感情推定
発表概要: 文に含まれる感情を推定することは対話システムやカスタマー分析への応用に必要な技術である。 従来、感情を推定する際にはSVMなどの分類器を用いて、単純に多値分類問題として取り組まれてきた。 本発表では、感情の対を一つの軸と見なして分類を行う新たな手法を述べる。
大串正矢 1151023: M, 2回目発表 中村哲, 松本裕治, 戸田智基, グラムニュービッグ, サクリアニサクティ
title: A Consideration of the feature-rich of Jointly Optimization in Speech Translation Systemns
abstract: ST systems consists of three components: an automatic speech recognizer (ASR) and a machine translator (MT), text to speech (TTS). In these systems, speech recognition and machine translation are conventionally optimized individually. Zhang et al. (2004) proved the effectiveness of the method to optimize speech recognition and machine translation using translation evaluation measures. However there are a limited number of features that can be handled in the optimization technique, it is not scalable to lager feature sets.In this research we inspected the effect of pairwise ranking optimization (PRO) an optimization technique highly scalable optimization techinique in the jointly optimization framework. In addition, we achieve improvements in the translation accuracy by adding rich features that express the frenquency of each word in the recognition candidate.
language of the presentation: *** Japanese ***
発表題目: 音声翻訳システムの同時最適化における拡張性の検討
発表概要:音声翻訳システムは音声認識(ASR)、機械翻訳(MT)、テキスト音声合成の3つのパートで構成されており、従来は音声認識と機械翻訳を個々に最適化する。しかし音声認識と機械翻訳の両者を同じ翻訳の尺度で最適化する手法の有効性が証明されている(Zhang et al.,2004)。現在までの最適化手法では扱える素性の数に限りがあり、拡張性に優れていない。そこで本研究では拡張性の高い最適化手法である対ランキング最適化(PRO)(Hopkins et al.,2011)を同時最適化に適用し、その効果を検証した。また、扱う素性に認識候補における各単語の頻度を加えることにより、フィラーなどによる翻訳誤りの改善を図ることとした。
小木曽智信 1161018: D, 中間発表 松本裕治, 中村哲, 新保仁, 小町守
title: Morphological analysis for historical Japanese texts
abstract: Since usual Japanese morphological analysis has aimed to analyze contemporary Japanese, it has been difficult to analyze historical Japanese texts such as classical Japanese. However, analysis of various old texts is needed in order to construct Japanese diachronic corpus that is currently being planned at NINJAL. In this study, I developed new dictionaries for MeCab by creating training corpus of old Japanese and preparing lexical entries based on UniDic. Analysis of Early Middle Japanese and Modern Literary Japanese has become possible by these dictionaries.
language of the presentation: Japanese
発表題目: 歴史的資料を対象とした日本語形態素解析の実現
発表概要: 従来の日本語形態素解析は専ら現代語を対象としてきたため,古典語等の歴史的な日本語テキストを解析することは困難であった。しかし国立国語研究所で準備中の通時コーパスの構築のためには過去の多様なテキストを解析する必要がある。本研究では,UniDicをベースとして辞書項目を整備し古語の学習用コーパスを作成することにより,新たなMeCab用形態素解析辞書を開発した。これにより中古和文・近代文語等の解析が可能となった。

会場: L2

司会: 武富 貴史
中村勇貴 1151077: M, 2回目発表 安本慶一, 加藤博一, 松本健一, 玉井森彦
title: Daily Life Support Method using AR Effect based on History of User Behavior
abstract: We propose a support system for everyday life based on the user's behavior log in the indoor environment. In general, habitual behavior is difficult to remain in the memory. There are some existing studies that support daily activity or improving work efficiency. Also, they do not discuss on intuitive user interfaces that alleviate users frustration during daily activities. In this research, we propose a user's history management technique which links each object/sub-space in the target indoor space to the user behavior log. Then, we propose a method for reproducing AR-based visual effect on the object/sub-space which naturally navigates the user to do the intended work. We have implemented a prototype of the proposed method and conducted subjective survey to evaluate the effectiveness of the AR-based visual effects.
language of the presentation: Japanese
発表題目: ユーザの行動履歴に基づくAR効果を用いた日常生活支援システム
発表概要: 本発表では,屋内環境における掃除等の日常生活を支援するシステムを提案する.一般的に,習慣的な行動はいつ行なったかなどを意識することがなく記憶に残りにくい.この問題に対して,作業支援や作業効率の改善を狙う研究ががいくつか存在するが,ある特定の行動・目的に対する支援が対象であるため,汎用性に欠けるという課題がある.また,ユーザのシステム使用時の煩わしさを軽減するような,直観性に優れた出力については議論されていない.本研究では,これらの課題に対し,ユーザの行動履歴を家具や空間に紐付けして履歴の管理手法を提案する.また,拡張現実感技術を用いて,空間を装飾するという直観的な出力を用いることで,自然にユーザの意識を誘導するような支援方法を提案する.また本発表では,提案手法のプロトタイプによる実験結果を報告する.
勝山貴史 1151037: M, 2回目発表 小笠原司, 加藤博一, 高松淳, 竹村憲太郎
title: Generating semantic 3D map using human behavior
abstract: To realization of intelligent service, semantic map, which includes variety of information, is required. In previous studies, semantic maps have been constructed using data obtained by sensors such as camera and LIDAR. Recently some researchers tried to correlate human behavior with an environmental map. However, previous researches did not consider Point Cloud, so I will propose the behavior-based semantic 3D map in this presentation.
language of the presentation: Japanese
足立麻衣子 1151006: M, 2回目発表 萩田紀博, 小笠原司, 神原誠之
title: Self-sacrificial Robot -A beatable robot for management of user’s emotional behavior-
abstract: The aim of this research is development of an interaction robot which can be beaten by a user for management of user’s emotional behavior. Though the beating is often recognized as a negative action, the beating is one of important action for venting of user’s emotional behavior. For example, we sometimes beat something to manage a stress, but beating others is unacceptable because of social moral and position. We try to resolve two problems to realize a robot which can be beaten by a user without hesitating. First problem is how to beat a robot without hesitating. To overcome this problem, we investigate physical and mental resistances of the user against the action of beating a robot. Second problem is beating action continuation. A robot behavior which keeps the robot being beaten by user continuously will be designed from results of case study.
language of the presentation: Japanese
高砂雄介 1151124: M, 2回目発表 萩田紀博, 小笠原司, 神原誠之
title: Collaborative cleaning with a human using an interactive communication robot
abstract: A goal of my research is development of an interactive cleaning robot which cooperated with a human. Though some types of cleaning robots have been developed as consumer product, these robots have some problems. The typical problems are that user cannot know area which has been cleaned and robot cannot clean if obstacles exist in cleaning area. We overcome these problems by collaborating with human using an interactive cleaning robot in ambient intelligence platform in which positions of the users and the robot can be measured in real time. The cooperative work of human and the robot is carried out by using ambient map which indicates the environment status including the human, objects and the robot with sensing information.
language of the presentation: Japanese