| 石倉 明生 | 1451007: M, 2回目発表 | 笠原 正治，安本 慶一，笹部 昌弘，川原 純 |
|
title: Construction of pedestrian flow simulation models
abstract: When we consider the effect of the evacuation guidance system by simulation, it is important to estimate positions from which evacuees start to head to shelter because the distribution of evacuees on roads affects the result of the simulation. Therefore, in order to provide initial positions of evacuees and the number of evacuees at the start of the simulation properly, I am constructing and implementing some models about how people move and developing methods how to estimate population distribution in their daily lives. In this research, using actual data of movement information, I considered whether the existing walking movement model fits the data by maximum likelihood estimation. In this presentation, I report the investigation result. I also describe the outlook on constructing walking movement models I propose and estimating population distribution in their daily lives by using the data. language of the presentation: Japanese 発表題目: 歩行者流動シミュレーションモデルの構築・実装 発表概要: 災害発生時の避難者の避難誘導方式をシミュレーションによって検討する際に，避難者の避難開始位置の推定は重要となる．本研究は，歩行移動モデルを構築するとともに対象とするエリアの平常時の人口分布を推定することで，災害発生時の避難者の位置や人数を推定することを目的としている．人の移動情報の実データを使用し，実データが既存の歩行移動モデルに当てはまるかの検討を最尤法によって行った．本発表では，検討結果を報告するとともに，実データを使用した新たに構築する歩行移動モデルや平常時の人口分布推定の展望について述べる． | ||
| 糸井 純暉 | 1451009: M, 2回目発表 | 笠原 正治，安本 慶一，笹部 昌弘，川原 純 |
|
title:
Design, implementation, and evaluation of an offline evacuation guiding application based on cooperation between evacuees and their mobile terminals.
abstract: There is an existing work which proposes automatic estimation of blocked road segments and evacuation guiding based on cooperation between evacuees and their mobile terminals so that the evacuees can quickly evacuate to safe places. To realize this scheme in actual systems, we have to implement a mobile application which has the following functions: estimation of blocked road segments, search of routes to safe places, presentation of various information, and sharing information with other terminals. In this study, we design and implement an offline version of the mobile application, which excludes the communication function. Through experiments, we evaluate the practicality and effectiveness of the mobile application. language of the presentation: Japanese 発表題目: 避難者と携帯端末の連携に基づくオフライン型避難誘導アプリケーションの設計，実装，評価 発表概要: 災害発生時に避難者を安全に素早く避難所に誘導することを目的として，避難者と携帯端末との連携を用いて通行不能箇所の推定と避難誘導を自動化する方式が 提案されている．実環境でこの方式を実現するためには，通行不能箇所の推定，避難所までの経路検索，各種情報の提示，他端末との情報共有，といった機能を携帯端末上のアプリケーションとして実装する必要がある．本研究では，特に，通信機能を除く，オフライン動作可能な機能を設計，実装する．さらに，実証実験を通して，実装アプリケーションの実用性と有効性を評価する． | ||
| 大槻 紘平 | 1451025: M, 2回目発表 | 笠原 正治，安本 慶一，笹部 昌弘，川原 純 |
|
title: Detection of outlier trajectories in evacuation guiding system based on cooperation between evacuees and their mobile terminals.
abstract: Recently, the evacuation guiding system using trajectory information (consisting of a sequence of the GPS locations) collected from evacuees’ mobile terminals has been developed (Komatsu et al., 2015). This system estimates impassable roads from the difference between the evacuation route that the mobile terminal which an evacuee owns has presented (recommended route) and the route that the evacuee has been actually through (movement trajectory). At this time, the estimation method does not work correctly if there are evacuees who do not move according to the recommended routes. Then, I am considering an estimation method for automatically detecting trajectories that the system intends them not to pass through by using the real-time burst detection method proposed by Ebina et al., 2010. In this presentation, I describe results and discussion of experiments to verify the estimated accuracy of this methods and a prospect of my future research. language of the presentation: Japanese 発表題目: 避難者・携帯端末連携型避難誘導方式における異常軌跡の検出 発表概要: 近年、災害発生時に避難者の携帯端末から収集した移動軌跡情報（GPS位置データの列）を用いて避難誘導を行うシステムの研究が行われている（小松ら 2015）。このシステムでは、携帯端末が提示した避難経路（推薦経路）と実際の避難行動（移動軌跡）との差から通行不能な道路の推定を行う。このとき、推薦経路と異なる目的地へと移動する避難者が存在すると推定方式が正常に動作しなくなる。現在、リアルタイムなバースト検出手法（蛯名ら、2010）を用いてこのようなシステムが意図しない軌跡情報を自動的に検知するための推定方式について検討を進めている。本発表では、その推定精度を検証した実験の結果と考察、今後の研究に関する展望を述べる。 | ||
| 小松 展久 | 1451048: M, 2回目発表 | 笠原 正治，安本 慶一，笹部 昌弘，川原 純 |
|
title: Evacuation Guiding Scheme Based on Cooperation between Evacuees and Their Mobile Terminals
abstract: When disasters occur, evacuees have to evacuate to safe places quickly. With the proliferation of mobile terminals, we aim to achieve an evacuation guiding system using mobile terminals of evacuees. It is important to suppress evacuees' operations of mobile terminals as much as possible because they are under stressful and/or dangerous situation during their evacuation. In this research, we propose a scheme which can automatically estimate blocked road segments and guide evacuees by taking into account the gap between routes recommended by the system and actual trajectories of the evacuees. Through simulation experiments, we clarify the effectiveness of the proposed scheme in terms of evacuation time and evacuation ratio. language of the presentation: Japanese 発表題目: 避難者・携帯端末連携型避難誘導方式 発表概要: 災害発生後，被災者は迅速に安全な避難先へと移動する必要がある．本研究では，携帯端末の高い普及率を考慮し，避難者自身の保有する携帯端末を用いた避難誘導の実現を目指す．特に，災害発生直後の避難者の緊張状態と被災地の危険性を考慮し，避難者の携帯端末に対する操作はできる限り少なくできることが望ましい．本研究ではシステムが避難者に提示した経路に対して， 実際に避難者が移動した軌跡との差に着目することで，通行不能箇所の把握と避難誘導の自動化を可能とする方式を提案する．シミュレーション評価により，提案方式による避難時間と避難完了率の改善度合いを明らかにする． | ||
| 濱口 拓男 | 1461008: D, 中間発表 | 松本 裕治，中村 哲，新保 仁，進藤 裕之 |
|
title:CFG-based Neural Networks and extend method.
abstract: Deep Neural networks became new standards and are being delevopling in many Domains. In NLP, there are encode-decode models composed of encoder and decoder. Encoder converts a label sequence to a vector and decoder generates a label sequence from a vector. We focus on these models. First, we propose CFG-based Neural Netword. Standard encode-decode model uses Recurrent Neural Network. Like HMM, Recurrent Neural Network has linear structure. But deep layer causes many problems, e.g. vanishing/exploding gradient. We reconsider how decoder works and extract two essential points. From this viewpoint,the difference of linear structure and tree structure is a difference of grammar. Second, we propose a kind of pretraining method. Many Neural Networks in NLP can't be applied ordinary pretraining methods, because these Neural Networks are defined recurrently/recursively. To overcome this problem, we take a curriculum learning approach. In curriculum learning,we train an easy task and train a more difficult task using same parameter. We assume a simple hypothesis an easy task needs fewer parameter. In Neural Networks,we think lower dimension as fewer parameter. But this assumption is not preferable for curriculum learning because easy task's dimension and difficult task's dimension are not same. We develop a method extending lower dimension to higher dimension. Finally, I summarize these models. language of the presentation: Japanese | ||
| 麻生 栄樹 | 1451005: M, 2回目発表 | 松本 裕治，中村 哲，新保 仁，進藤 裕之 |
|
title: *** Automatic Extraction of Structured Abstract from Scientific Articles ***
abstract: *** *** language of the presentation: *** Japanese *** 発表題目: *** 科学技術論文からのstructured Abstractの自動生成 *** 発表概要: *** *** | ||
| 有山 祐平 | 1451006: M, 2回目発表 | 松本 裕治，中村 哲，新保 仁，進藤 裕之 |
|
title: Neural Networks for Entity Disambiguation
abstract: Entity disambiguation is a fundamental task in the field of natural language processing.Entity disambiguation in natural language text maps mentions of ambiguous names onto entities registered in a knowlege base such as Wikipedia.When this technology is developed improvement of meaning understanding of sentence can be expected.In this presentation, I will introduce conventional neural network approach, and then show my research plan. language of the presentation: Japanese 発表題目: ニューラルネットを用いた固有名詞の曖昧性解消 発表概要: 自然言語処理の分野において固有名詞の曖昧性解消は，重要なタスクとされています．自然言語のテキストにおける固有名詞の曖昧性解消とは，曖昧な固有名詞をWikipediaのような知識ベースに登録されているURIを紐づけることである．このタスクが発展するとユーザーの文章に対する意味理解の向上が期待できる．本発表では，ニューラルネットを用いた手法を紹介し，現状の進捗状況について報告するとともに，今後の予定について述べる． | ||
| 大西 真輝 | 1451026: M, 2回目発表 | 松本 裕治，中村 哲，新保 仁，進藤 裕之 |
|
title: Complex Coreference Resolution using Big Dependency Knowledge
abstract: Coreference resolution is the task of finding all expressions that refer to the same entity in a text. In this study, we examine complex case of them which traditional linguistic constraints on coreference as well as commonly-used resolution heuristics are not useful. We use a knowledge-rich approach to this task by a lot of dependency knowledge from annotated Gigaword. language of the presentation: Japanese 発表題目: 大規模依存構造知識を用いた「複雑な共参照問題」の解析 発表概要: 共参照解析とは，文書中の表現を現実世界の同一な実体ごとに分類するタスクである．本研究では特に，自然言語処理で使用される伝統的な技法では解くことが難しいとされている，「複雑な共参照問題」について取り扱う．我々は，メタデータが付与された大規模なGigawordコーパスを，依存構造知識として利用することでこのタスクに取りんだ． | ||