日時: 9月22日(木)2限(11:00-12:30)

会場: L1

鎌倉 まな D, 中間発表 知能コミュニケーション 中村 哲, 荒牧 英治, 鳥澤 健太郎, 飯田 龍
title: An Open-domain Counterargument Discovery Method Using a Large-Scale Event Causality Database Automatically Constructed from the Web
abstract: Argumentation being an important medium for exchanging opinions with others and making decisions, there have been studies on argumentative dialogue systems and methods for retrieving or generating counterarguments. Systems and methods proposed in such previous research only use small corpora manually created for research purposes, or political texts extracted from argument forums on the Web. By limiting the data this way, the quality and range of the counterarguments that these systems or methods produced has significantly deteriorated. In our own research, we propose a method that provides counterarguments to a wide range of input sentences, using a large set of texts representing event causalities that were automatically extracted from the Web using deep learning technologies. In our method, given an input sentence, we first retrieve candidates for counter-argumentation from a large-scale event causality database. Then, we use deep-learning technologies to identify counterargument candidates that describe undesirable situations happening as a consequence of the input sentence content, and provide those as actual counterarguments to the input sentence. As a preliminary study, we conducted experiments by selecting counterarguments using existing neural-based methods for recognizing entailment relations and trouble expressions, and found that we could this way successfully provide counterarguments against a wide range of input sentences. To further improve the quality of such counterargument acquisition, we propose a novel method that uses a single neural model to recognize relations between an input sentence and both parts (cause and effect) of a causal relation. We evaluated this method using human evaluation results, and finally discuss future work.
language of the presentation: Japanese
発表題目: 多様な分野における議論に適用可能な、Webから抽出した大規模因果関係データベースを用いた反論提示手法
発表概要: 一般に「議論」とは、他者と意見交換を行い意思決定を行う重要な手段であり、これまで、ユーザと議論する対話システムや、入力文に対する反論を検索・生成する手法が研究されてきた。既存の反論の検索・生成手法の研究では、研究用に人手で作成された小規模なコーパスか、Web上の議論フォーラムから抽出した政治・社会問題等の領域のテキストを使用するため、領域を限定しない多様な議論に関しては高品質な反論を出力することができない。本研究では、事前にWebから深層学習を用いて領域を問わず自動抽出した大規模因果関係知識を用いて、多様な入力文に対する反論を出力する手法を提案する。提案手法では、入力文に関する反論の候補を大規模因果関係データベースから検索し、入力文に書かれた内容が起こると付随して起こり得る望ましくない事態・動作・状態を含む反論候補を深層学習を用いて自動的に特定し、それを入力文に対する反論として出力する。予備調査として、まず、深層学習ベースの既存の含意関係認識とトラブル表現認識を組み合わせて反論を選択する実験を行い、人手評価により、広範な話題に対して反論を出力できる可能性を確認した。さらに反論の精度を高めるため、入力文と因果関係の原因部分と帰結部分の3文の関係を単一のニューラルモデルで認識する新しい手法を提案する。これらの手法により出力した反論の人手評価結果を報告し、今後対処すべき課題について論じる。
ZHOU YANGYANG M, 2回目発表 ソーシャル・コンピューティング 荒牧 英治, 渡辺 太郎, 若宮 翔子, Liew Kong Meng, She Wang Jou

title: Analyzing the relationship between weather and music preference

abstract: Previous research has considered individual or personal level influences of music preference extensively, but there has been less research considering environmental variables like weather patterns. In this paper, we aim to find whether weather patterns are related to music preference at the city level. Music preference was estimated through Top25 tracks from 106 cities, and corresponding weather data was collected from Apple Music and OpneWeather, respectively. We then compared weather variables to acoustic features (obtained from Spotify) for these tracks by comparing their K-means clusters (Adjust Rand Index) for both sets of data, but found no relationship. Following which, we analyzed from a microscopic perspective and found that mean daily temperature was negatively correlated with the danceability, energy, and tempo features, but positively correlated with loudness. In conclusion, we found only a limited relationship between weather and music preference at the city level: a significant relationship was observed for acoustic features pertaining to emotion and daily temperature. These findings complement the existing literature on the relationship between environmental variables and music preference, clarifying the relationship between ecology and music preference at the city level, and may potentially have implications for music recommendation algorithm development. 

language of the presentation: English

FENG XINCAN M, 2回目発表 自然言語処理学 渡辺 太郎, 湯上 伸弘, 鄭 育昌, 上垣外 英剛
title: *** Sharing Parameter by Conjugation for Knowledge Graph Embeddings in Complex Space ***
abstract: *** A Knowledge Graph (KG) is the directed graphical representation of entities and relations in the real world. KG can be applied in diverse Natural Language Processing (NLP) tasks where knowledge is required. The need to scale up and complete KG automatically yields Knowledge Graph Embedding (KGE), a shallow machine learning model that is suffering from memory and training time consumption issues. To mitigate the computational load, we propose a parameter-sharing method, i.e., using conjugate parameters for complex numbers employed in KGE models. Our method improves memory efficiency by 2x in relation embedding while achieving comparable performance to the state-of-the-art non-conjugate models, with faster, or at least comparable, training time. We demonstrated the generalizability of our method on two best-performing KGE models $5^{\bigstar}\mathrm{E}$ \cite{Nay21} and $\mathrm{ComplEx}$ \cite{Tro16} on five benchmark datasets. ***
language of the presentation: *** English ***

会場: L2

司会: Md.Delwar HOSSAIN
瀨古 巽 M, 2回目発表 サイバーレジリエンス構成学 門林 雄基, 笠原 正治, 岡田 実, 妙中 雄三
title: Conducting exercises on CISO work
abstract: The CISO is responsible for overseeing cybersecurity in a company and is an extremely important position in terms of comprehensive implementation of both technical and management/business roles. However, the appointment rate of CISOs in Japan is lower than in Europe and the United States. This is partly due to the fact that the roles and responsibilities of the CISO are not clearly defined. Therefore, we designed an exercise that enables role-playing of CISO duties using instructional design methods. The results of this exercise, which was conducted for master's students at several universities, are reported.
language of the presentation: Japanese
発表題目: CISO業務を題材とした演習の実施
発表概要: CISOは企業におけるサイバーセキュリティを統括する責任者であり、技術的な役割と経営・事業的な役割の両方を包括的に実施する点で極めて重要な役職となっている。しかし、国内における専任のCISOの任命率は欧米に比べ低い。これは、CISOの役割や責任が明確になっていないことも一因としてあげられる。そこで、CISO業務をロールプレイできる演習をインストラクショナルデザインの手法に乗っ取り設計した。この演習を複数大学の修士学生に対して実施したのでその結果を報告する。
小寺 智仁 M, 2回目発表 ロボットラーニング 松原 崇充, 安本 慶一, 鶴峯 義久
title: Development of Deep Reinforcement Learning Algorithms Robust to Function Approximation Errors in Spiking Neural Networks
abstract: In recent years, we have seen great interest in the use of deep reinforcement learning to automate edge robots. Since the power consumption of edge robots is finite, attention has focused on neurochips that consume extremely low power during computation compared to conventional GPUs and CPUs, and spiking neural networks, which are computational algorithms for this purpose. However, there is no learning method for spiking neural networks that can achieve realistic performance for real problems. Therefore, we focused on a method that operates by converting an artificial neural network to a spiking neural network, which is currently the most effective learning method. However, this learning method has a problem that the conversion causes function approximation errors, and in deep reinforcement learning, where regression accuracy is important, the conversion degrades the learning performance. In this study, we propose a deep reinforcement learning algorithm method that is robust to function approximation errors when converting to a spiking neural network. The proposed method solves the problem by using GIO, which updates the network composed of low bit widths by STE and enlarges the difference in value function values of each action during learning. As a result, we were able to learn measures with high performance on low-bit networks, and these networks have reduced errors when converted to spiking neural networks.
language of the presentation: Japanese
発表題目: スパイキングニューラルネットワークの関数近似誤差に頑健な深層強化学習アルゴリズムの開発
発表概要: 近年、深層強化学習を用いたエッジロボットの自動運転に大きな期待が寄せられている。エッジロボットの電力は有限であることから従来のGPUやCPUに比べて演算時の消費電力が極めて低いニューロチップと、そのための計算アルゴリズムであるスパイキングニューラルネットワークが注目されている。ただし、このスパイキングニューラルネットワークは実問題に対して現実的な性能を発揮できる学習方法が存在しない。そこで、現状最も有効な学習手段である従来の人工ニューラルネットワークをスパイキングニューラルネットワークに変換することで動作させる手法に着目した。しかし、この学習手法は変換によって関数の近似誤差が発生し、回帰精度が重要である深層強化学習では、変換によって学習性能が低下するという問題がある。そこで本研究では、スパイキングニューラルネットワークへの変換時の関数近似誤差に頑健な深層強化学習アルゴリズム手法を提案する。提案手法では、STEにより低ビット幅で構成されたネットワークを更新し、各行動の価値関数値の差を学習時に拡大させるGIOを用いることで問題を解決する。結果として、低ビットのネットワークでも性能が高い方策を学習することができ、これらのネットワークはSNNへの変換時の誤差が減少した。
鉢峰 拓海 M, 2回目発表 ロボットラーニング 松原 崇充, 和田 隆広, 佐々木 光
Cutting Surface Model-Based Reinforcement Learning for Object Shape Manipulation : Application to Rough Grinding
Removal processes are used to initial shape into a target shape by removing the material, such as cutting vegetables with a knife. During these processes, removal resistance is evolved between the object and the robot, depending on the processing conditions. Therefore, uncertainty accumulate during the shaping transition due to slippage and blurring. For this reason, we need to consider the shape uncertainty depending on the processing conditions. In this study, we propose a shape transition model that is divided into a geometric cutting model independent of processing conditions and a model which compensates for uncertainties caused by processing conditions. The effectiveness of the proposed method are confirmed by the rough grinding process in the simulator and the real robot, respectively.
Language of the presentation : Japanese