| 浅井 沙良 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 宮尾 知幸(MS), 吉野 幸一郎, Sakriani Sakti |
|
title: Building Persuasive Dialog System using Emotional Expressions
abstract: Emotional expressions are known to be effective in persuading others of your opinions. They are also expected to work well in persuasive dialog systems. However, conversation systems have difficulties in controlling emotional states, because typical dialog corpuses include utterances followed by only one pattern of responses. In our study, aiming to achieve an effective persuasion, we first build a response generation model to reflect emotions. With the results, we also introduce an augmentation of persuasive dialog corpuses and collection of voice data to express various emotional expressions. language of the presentation: Japanese 発表題目: 感情表現を行う説得対話システムの構築 発表概要: 感情表現は、説得において自分の意見を相手に受諾させる手段として効果的であることが知られている。ユーザを説得する対話システムにおいても感情表現の有用性が期待される。しかし、通常の対話コーパスでは発話に続く応答は一つのパターンしか存在しないため、対話システムで感情状態を制御させることは困難である。本研究では、感情を用いた効果的な説得を目指して、感情を反映する応答文の生成モデルを構築した。また、この結果を踏まえ、説得対話システムにより多様な感情表現を伝達させる方法として、コーパスの拡張と音声によるデータ収集についても紹介する。 | |||
| 石橋 陽一 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 宮尾 知幸(MS), 須藤 克仁 |
|
title: Reflection-based Word Attribute Transfer
abstract: We propose a word attribute transfer framework based on ``reflection'' to obtain a word vector with an inverted target attribute for a given word, in a word embedding space. Word embeddings based on Pointwise Mutual Information (PMI) represent analogic relations such as ``king'' - ``man'' + ``woman'' \approx ``queen''. Such a relation can be used for changing a certain attribute of a word, such as a transfer from ``king'' into ``queen'' by changing its attribute of gender. This attribute transfer can be performed by adding a difference vector ``woman'' - ``man'' to ``king'', when we have explicit knowledge of the gender of the given word ``king''. However, this knowledge cannot be developed for various words and attributes in practice. In the case of transferring ``queen'' into ``king'' in this analogy-based manner, we need to know ``queen'' is a female word and to choose to subtract the difference vector from ``queen''. In this work, we aim to transfer such a binary attribute based on an assumption that such a transfer mapping should become an identity mapping when we apply it twice. We introduce a framework based on reflection mapping that satisfies this property; ``queen'' should be transferred back into ``king'' with the same mapping as the transfer from ``king'' to ``queen''. Experimental results show that the proposed method is able to transfer word attributes of given words, with keeping words not having the target attributes unchanged. language of the presentation: Japanese | |||
| 髙橋 洸丞 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 宮尾 知幸(MS), 須藤 克仁 |
|
title: Cross-lingual Language Model Based Automatic Machine Translation Evaluation using Source Language Feature
abstract: We investigate an evaluation metric which uses source and reference sentences for evaluation. Many previous studies are based on a comparison between 1-to-1 paired reference and machine translated sentences. However, 1-to-1 paired comparison tends to miss out the cases where translations are correct but different from references in the surface. While this problem can be solved by introducing other reference sentences, which is callled multi-reference, preparing references is costly for the paired languages. Thus, our approach is to regard source senteces as pseudo reference senteces and utilize source information in evaluation. We investigated an evaluation model which predicts scores in regresion manner by using embbed input sentences with cross-ingual language model. language of the presentation: Japanese 発表題目: 言語横断な言語モデルによる原言語情報を活用した自動機械翻訳評価 発表概要: 本研究では原言語文と参照訳文から翻訳文がどれだけ正しいかを推定する自動評価手法を検討する. 既存の自動翻訳評価手法では, 1 対 1 の参照訳文と翻訳文のペアから翻訳文を評価する手法が主流だが, 1 対 1 ペアの比較では, 翻訳文が参照訳文と一見異なるが正しい訳文である際に評価性能が下がりやすい. こ の問題は, マルチリファレンスと呼ばれる参照訳文を複数用意することで解決可能だが, 各原言語文に対し て参照訳文を複数作成する必要がありコストが高い. そこで本研究では, 原言語文も参照訳文の一つとみな して, 原言語文を評価に用いることで前述の問題に対処する. 言語横断な言語モデルを用いて入力文を分散 表現とし, 最終的な評価値を多層パーセプトロンを通して出力する回帰モデルによる評価手法を検討した. | |||
| 恒松 和輝 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 宮尾 知幸(MS), Sakriani Sakti |
|
title: Listen, Predict, and Speak with Deep Learning
abstract: In human-machine interaction, recent technologies such as speech recognition (ASR) and speech synthesis (TTS) have enabled a machine to listen and speak like a human. But ASR only passively transcribes what was being said. TTS only passively read out loud what was being written. On the other hand, in human-human communication, human often predicts the sentence even the other person has not finished speaking the complete sentence. In this research, we aim to construct a system that can perform such tasks using deep learning. Text prediction was performed on various models, including the text-to-text and speech-to-text models. Although seq2seq is different from the reference sentence, you can see that the system is able to produce a complete natural sentence. In the future, we will perform a speech-to-speech prediction directly. language of the presentation: Japanese 発表題目: 入力音声に続く文章の予測 発表概要: 近年の技術の進歩により，音声認識は人間に近い非常に高い性能を示している．しかし，それらは与えられた音声をただ文字に起こすだけである．人間同士の対話では，発せられた言葉からその後に続く言葉を予測できることがある．本研究では，深層学習を用いてそのようなタスクを実行できるシステムの構築を目指す．様々なモデルを用いて，音声での予測と，文章のみでの予測を行なった．Sequence-to-Sequenceでは，参照文とは異なるが自然な文章を出力できることが分かった．今後は，出力を文章ではなく音声に変更して同様の実験を行う． | |||
| 今岡 一章 | M, 2回目発表 | ネットワークシステム学 | 岡田 実, 笠原 正治, 東野 武史, Dong Duong Thang |
|
title: An Investigation of Rainfall Forecasting Using GNSS Zenith Total Delay with Elevation Cut-off Angle
abstract: As one of the meteorological observation techniques, the Global Navigation Satellite System (GNSS) is known. This technique utilizes the fact the GNSS signal is delayed when it goes through the troposphere with high water vapor in rhe atmosphere, and causes positioning error. This extra delay time is treated as Zenith Total Delay (ZTD), and the ZTD is internally estimated on positioning calculation in the GNSS receiver. However, an increase of ZTD does not always indicate the occurrence of rainfall. In this presentation, we focus on the elevation cut-off angle, one of the setting of positioning, to prevent the false alarm in rainfall forecasting. This setting plays an important role for rainfall forecasting, because it determines the visible range in the viewpoint of the GNSS receiver for seeking rain clouds. To verify yhe possibility, we computed ZTD time series under the varying elevation cut-off angles and compared with rainfall events. These results show a possibility of distinguishing the increase of ZTD due to the occurrence of rainfall from the increase of ZTD due to other factors. language of the presentation: Japanese | |||
| 長濵 将太 | M, 2回目発表 | ディペンダブルシステム学 | 井上 美智子, 笠原 正治, 大下 福仁, 新谷 道広 |
|
title: Graph Exploration of Autonomous Mobile Robots with Limited Visibility
abstract: In this research, we investigate graph exploration algorithms for autonomous mobile robots. The robots have limited visibility, that is, they observe nodes within a certain fixed distance, and luminous, that is, they have light devices that can emit a color from a set of constant number of colors. We have studied ring exploration and finite grid exploration so far. For ring exploration, we considered the constraint that the visible distance is any constant of at least two and the number of colors of light devices is two. Under the constraint, we proved that 1) two robots are necessary and sufficient to achieve perpetual exploration and 2) three robots are necessary and sufficient to achieve terminating exploration, where perpetual exploration requires every robot to visit every node infinitely many times and terminating exploration requires robots to terminate after every node is visited by a robot at least one. These results show that the number of robots required for exploration can be reduced by extending their visibility compared to the previous work. We also showed that the proposed perpetual exploration algorithm for rings is universal, that is, the algorithm achieves perpetual exploration from any solvable initial configuration with two robots. On the other hand, we showed that no universal algorithm exists for terminating ring exploration with three robots. For finite grid exploration, we designed terminating exploration algorithms for grid under various constraints. This is the first work of finite grid exploration with limited visibility. language of the presentation: Japanese | |||
| 安見 嘉人 | M, 2回目発表 | ディペンダブルシステム学 | 井上 美智子, 笠原 正治, 大下 福仁, 新谷 道広 |
|
title: Uniform Partition in A Population Protocol Model
abstract: In recent years, autonomous distributed systems using low-performance devices have attracted attention. With such low-performance devices, existing algorithms of distributed systems cannot be applied because functions of devices are greatly limited. Hence, in this research, we use a population protocol model that is a model for such low-performance devices, and consider the space complexity to solve basic tasks in the model. In particular, we focus on the uniform k-partition problem as a basic task. The goal of the uniform k-partition problem is to divide a population into k groups of the same size (k is a positive integer). An application of the uniform k-partition problem is to execute multiple tasks simultaneously by assigning different tasks to each group. The main results of this research are the following. 1) We completely clarify the space complexity of the uniform 2-partition problem for each combination of four assumptions (existence of the base station, fairness, initial state of devices, and symmetry of the protocol). 2) We consider the space complexity of the uniform k-partition problem for each combination of those four assumptions. As a result, we clarify the space complexity for many combinations of assumptions. 3) For each combination of those four assumptions, we consider the space complexity of the uniform 2-partition problem in the model with new additional assumption that pairs of devices that can communicate are restricted. As a result, we clarify the space complexity for many combinations of assumptions. language of the presentation: Japanese | |||
| 桃井 雄資 | M, 2回目発表 | 大規模システム管理 | 笠原 正治, 伊藤 実, 笹部 昌弘 |
|
title: On Enumeration of Matching Solutions by Zero-suppressed Binary Decision Diagram
abstract: Making a good matching between two-layer vertices is one of the most important problems not only in mathematics but also in some practical situations like the Hospital/Residents problem. There are a huge number of candidates for matching solutions in a graph. Each vertex has its own wish list of vertices in the other side. If vertices are matched to dislikes or they feel they could get more appropriate partner, they may complain of their matching result. Stable matching is one of the solutions for their discontent, because no other pair likes each other more than matching pairs. Gale and Shaplay proposed an algorithm to obtain a stable matching solution in polynomial time. However, many instances with randomized wish lists have few stable matchings, and they are often biased for side-based satisfaction. In this research, we consider to permit a matching to have a specified number of blocking pairs and enumerate all possible matchings with at most k blocking pairs by using zero-suppressed binary decision diagram (ZDD), which is a data structure for representing a set family, in order to offer good choices in matching problems. First, we propose an algorithm to enumerate all stable matchings for a given complete wish list, and then compare it with an enumeration algorithm based on the satisfiability problem. Second, we regard a new algorithm to enumerate all matchings with at most k blocking pairs for an incomplete wish list. Finally, we suggest some experiments confirming the performance of our algorithms. language of the presentation: Japanese | |||