研究者詳細 - 嶋田　和孝

論文 - 嶋田　和孝

分割表示146 件中 61 - 80 件目／全件表示 >>

Sarcasm detection using RnN with relation vector 査読有り

Hiai S., Shimada K.

International Journal of Data Warehousing and Mining 15 ( 4 ) 66 - 78 2019年10月

担当区分：最終著者記述言語：英語掲載種別：研究論文（学術雑誌）

Copyright © 2019, IGI Global. Sarcasm detection has been treated as a task that classifies text as sarcastic or non-sarcastic. Sarcasm detection is a significant challenge for sentiment analysis because sarcasm involves a positive expression with a negative meaning. Surface information in text is commonly used as a classification feature. However, the authors must consider both surface and non-surface features. In this article, the authors focus on relation information between pairs of role expressions, such as “boss and staff,” and propose a sarcasm detection method based on surface and relation information. First, the authors extract role pairs from a corpus. Then, the authors construct a relation vector generated from these role pairs and incorporate the relation vector into a recurrent neural network model. The authors evaluated the proposed method by comparing it to previously proposed methods. The results demonstrate the effectiveness of introducing the relation vector to sarcasm detection.

DOI： 10.4018/IJDWM.2019100104

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85072636831&origin=inward

Timing Prediction of Facilitating Utterance in Multi-party Conversation 査読有り

Tomonobu Sembokuya, Kazutaka Shimada

Proceedings of PACLING 2019 2019年10月

Construction of a liar corpus and detection of lying situations 査読有り

Takabatake S., Shimada K., Saitoh T.

Proceedings - 2018 Joint 10th International Conference on Soft Computing and Intelligent Systems and 19th International Symposium on Advanced Intelligent Systems, SCIS-ISIS 2018 971 - 976 2019年05月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© 2018 IEEE. Recognizing lies in human interaction is one of the most important and challenging tasks in artificial intelligence. For example, it leads to the development of a human-computer interaction system that detects a liar in a conversation. In another example, some systems in the future might occasionally need to tell a lie for not hurting user's feelings, namely a white lie for trouble-free conversations. In this paper, we construct a video corpus for developing a system that detects lies in conversations. The corpus consists of 540 question-answer pairs by 9 persons; 270 true answers and 270 lies. We manually analyze the images in the corpus for recognizing lies. In addition, we apply machine learning techniques, support vector machines with local binary pattern features and geometric features, to our corpus. We discuss the results of the analysis of the corpus and the machine learning approaches to our corpus.

DOI： 10.1109/SCIS-ISIS.2018.00161

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85067133291&origin=inward

Annotation and analysis of extractive summaries for the kyutech corpus 査読有り

Yamamura T., Shimada K.

LREC 2018 - 11th International Conference on Language Resources and Evaluation 3216 - 3220 2019年01月

記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© LREC 2018 - 11th International Conference on Language Resources and Evaluation. All rights reserved. Summarization of multi-party conversation is one of the important tasks in natural language processing. For conversation summarization tasks, corpora have an important role to analyze characteristics of conversations and to construct a method for summary generation. We are developing a freely available Japanese conversation corpus for a decision-making task. We call it the Kyutech corpus. The current version of the Kyutech corpus contains topic tags of each utterance and reference summaries of each conversation. In this paper, we explain an annotation task of extractive summaries. In the annotation task, we annotate an importance tag for each utterance and link utterances with sentences in reference summaries that already exist in the Kyutech corpus. By using the annotated extractive summaries, we can evaluate extractive summarization methods on the Kyutech corpus. In the experiment, we compare some methods based on machine learning techniques with some features.

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85059885448&origin=inward

Automatic Training Data Construction and Extractive Supervised Summarization for NTCIR-14 QA Lab-PoliInfo 査読有り

Hiai S., Otani Y., Yamamura T., Shimada K.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11966 LNCS 97 - 109 2019年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© 2019, Springer Nature Switzerland AG. On the summarization task at NTCIR-14 QA Lab-PoliInfo, participants of the task need to generate a summary corresponding to an assembly member speech in assembly minutes within the limit length. Our method extracts important sentences to summarize an assembly member speech in the minutes. Our method applies a machine learning model to predict the important sentences. However, the given assembly minutes’ data do not contain information about the importance of the sentences. As a result, we cannot directly utilize machine learning techniques for the task. Therefore, we construct training data for the importance prediction model using a word similarity between sentences in a speech and those in the summary. In addition, we apply the sentence reduction process. In the process, we consider characteristics of summaries of assembly minutes to avoid removal of important words in extracted sentences. On the evaluation, all the scores by our supervised method with the constructed data outperformed unsupervised and supervised baseline methods. The result shows the effectiveness of our method.

DOI： 10.1007/978-3-030-36805-0_8

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85076978347&origin=inward

Trivia Score and Ranking Estimation Using Support Vector Regression and RankNet 査読有り

Kazuya Niina, Kazutaka Shimada

Proceedings of PACLIC32 2018年12月

Dialogue Act Annotation and Identification in a Japanese Multi-party Conversation Corpus 査読有り

Takashi Yamamura, Masato Hino, Kazutaka Shimada

Proceedings of Asia-Pacific Corpus Linguistics Conferenc (APCLC2018) 2018年09月

スポーツ要約生成におけるテンプレート型手法とニューラル型手法の提案と比較査読有り

田川裕輝, 嶋田和孝

自然言語処理 ( 一般社団法人　言語処理学会 ) 25 ( 4 ) 357 - 391 2018年09月

担当区分：最終著者,　責任著者記述言語：日本語掲載種別：研究論文（学術雑誌）

<p>本研究では，日本で人気のある野球に着目し，Play-by-play データからイニングの要約文の生成に取り組む．Web 上では多くの野球に関する速報が配信されている．戦評は試合終了後にのみ更新され，“待望の先制点を挙げる”のような試合の状況をユーザに伝えるフレーズ（本論文では Game-changing Phrase; GP と呼ぶ）が含まれているのが特徴であり，読み手は試合の状況を簡単に知ることができる．このような特徴を踏まえ，任意の打席に対して，GP を含む要約文を生成することは，試合終了後だけでなく，リアルタイムで試合の状況を知りたい場合などに非常に有益であるといえる．そこで，本研究では Play-by-play データから GP を含む要約文の生成に取り組む．また，要約生成手法としてテンプレート型文生成手法と Encoder-Decoder モデルを利用した手法の 2 つを提案する．</p>

DOI： 10.5715/jnlp.25.357

CiNii Article

CiNii Research

その他リンク： https://www.jstage.jst.go.jp/article/jnlp/25/4/25_357/_pdf

Image-based Attention Level Estimation of Interaction Scene by Head Pose and Gaze Information 査読有り

Komiya R., Saitoh T., Shimada K.

Proceedings - 17th IEEE/ACIS International Conference on Computer and Information Science, ICIS 2018 497 - 501 2018年09月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© 2018 IEEE. In many situations such as chat and discussion, we interact with other people, and interaction analysis is useful for improving communication skills and evaluating a learning effect of education. Although the audio information is possible to analyze the speaker behavior, it is difficult to analyze the non-speaker. On the other hand, the image information is possible to estimate the inside information, such as the concentration level and understanding level of the interaction of the non-speaker not only the speaker by using nodding, face direction, facial expression, and gaze. This paper proposes an image-based method to automatically estimate the attention level of participants for the interaction scene. We collected simulated interaction scenes by using the omnidirectional camera, and applied the proposed method. As a result, we confirmed the effectiveness of the proposed method.

DOI： 10.1109/ICIS.2018.8466462

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85055721201&origin=inward

Discussion map with an assistant function for decision-making: a tool supporting consensus-building 査読有り

Ryunosuke Kirikihira, Kazutaka Shimada

Proceedings of the tenth International Conference on Collaboration Technologies (CollabTech 2018) 2018年09月

Analysis of facilitators' behaviors in multi-party conversations for constructing a digital facilitator system 査読有り

Tsukasa Shiota, Takashi Yamamura,Kazutaka Shimada

Proceedings of the tenth International Conference on Collaboration Technologies (CollabTech 2018) 2018年09月

スポーツ要約生成におけるテンプレート型手法とニューラル型手法の提案と比較査読有り

田川裕輝, 嶋田和孝

自然言語処理 25 ( 4 ) 2018年09月

Image-based Attention Level Estimation of Interaction Scene by Head Pose and Gaze Information 査読有り

Rinko Komiya, Takeshi Saitoh, Kazutaka Shimada

Proceedings of the17th IEEE/ACIS International Conference on Computer and Information Science (ICIS 2018) 497 - 501 2018年06月

Simple and sophisticated inning summary generation based on encoder-decoder model and transfer learning 査読有り

Tagawa Y., Shimada K.

Proceedings of the 2017 International Conference on Asian Language Processing, IALP 2017 2018-January 252 - 255 2018年02月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© 2017 IEEE. This paper describes an inning summarization method for a baseball game by using an encoder-decoder model. Each inning in a baseball game contains some events, such as hits, strikeouts, homeruns and scoring. Simplified description of the events leads to the improvement of readability of the inning information. Our method learns a relation between play-by-play data in each inning and inning reports. We also incorporate sophisticated expressions acquired from game summaries with the model. We call them Game-changing Phrase, GP. One problem in our task is the size of training data for the learning. To solve this problem, we apply a transfer learning approach into our method. In the experiment, we evaluate the effectiveness of our method with the transfer learning.

DOI： 10.1109/IALP.2017.8300591

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85046629402&origin=inward

Stock market prediction using keywords from expert articles 査読有り

Ichinose K., Shimada K.

Advances in Intelligent Systems and Computing 700 409 - 417 2018年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© 2018, Springer International Publishing AG. The market analysis is one of the important tasks for text mining. In this situation, Web news has an important role to predict stock prices. In this paper, we propose a method to predict the Nikkei Stock Average, which is one of the most important stock market indexes. We extract viewpoints from experts’ articles for analyzing Web news. The extracted words are index words in the vector space of a machine learning technique. We also incorporate word embedding and bootstrap approaches into our method. It predicts “UP” or “DOWN” of the next day by using the articles of a day. We also evaluate our method with not only one-day prediction but also simulated trading. The experimental result shows that index words based on expert articles were effective for both one-day prediction and simulated trading.

DOI： 10.1007/978-3-319-72550-5_39

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85041546156&origin=inward

Sports game summarization based on sub-events and game-changing phrases 査読有り

Tagawa Y., Shimada K.

Studies in Computational Intelligence 742 65 - 80 2018年01月

記述言語：英語掲載種別：研究論文（学術雑誌）

© Springer International Publishing AG 2018. Microblogs are one of the most important resources for natural language processing. This paper describes a summarization task of sports events on Twitter. We focus on an abstractive approach based on sub-events in the sports event. Abstractive summaries usually are better than summaries generated by extractive approaches in terms of readability. Furthermore, our method can incorporate sophisticated phrases that explain the scene. First, our method detects burst situations in which many users post tweets when a sub-event in a game occurs. Tweets in the burst situations are the inputs of our method. Next, it extracts sub-event elements (SEEs) that contain actions in a game, such as “Player A made a pass to Player B” and “Player B made a shot on goal.” Then, it identifies the optimal order of the extracted SEEs by using a scoring method. Finally, it generates an abstractive summary on the basis of the ordered SEEs, such as “Playler B made a shot on goal from the Player A’s pass.” In addition, it adds game-changing phrases into the abstractive summary by some rules. In the experiment, we show the effectiveness of our method as compared with related work based on an extractive approach.

DOI： 10.1007/978-3-319-70636-8_5

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85041708730&origin=inward

Sarcasm detection using features based on indicator and roles 査読有り

Hiai S., Shimada K.

Advances in Intelligent Systems and Computing 700 418 - 428 2018年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© 2018, Springer International Publishing AG. Sarcasm is a non-literalistic expression and presents a negative meaning with positive expressions. Sarcasm detection is a significant challenge for sentiment analysis which is to analyze documents with opinions. In this study, we propose a method of sarcasm detection on Twitter. We focus on two kinds of feature words. One is words modified by the indicator “ ”. The other is words expressing a role. First, we extract these words from tweets. Next, our method uses the lists of these words for a machine learning approach to detect sarcastic tweets. The lists of extracted words are used as features in our method. In the experiment, we compare our method with a baseline based on the features in previous studies. The experimental result shows the effectiveness of our method.

DOI： 10.1007/978-3-319-72550-5_40

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85041536268&origin=inward

Trivia score and ranking estimation using support vector regression and ranknet 査読有り

Niina K., Shimada K.

Proceedings of the 32nd Pacific Asia Conference on Language, Information and Computation, PACLIC 2018 263 - 270 2018年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

© 2018 by the authors. Dialogue systems have been increasingly important these days. In particular, non-taskoriented dialogue systems are studied becauseof the success of neural network approachessuch as seq2seq models. However, these models tend to generate simple responses such as"yes" and "ok." To construct a dialogue system that holds users' attention continuously,we need to generate utterances that capturethe interest of the user. In this paper, wepropose a method to extract trivia sentencesfor the purpose. Trivia information perhapsadds a surprise to users. Therefore, capturing trivia information is beneficial for dialogue systems. We estimate a trivia scoreof a sentence by using machine learning approaches, Support Vector Regression (SVR)and RankNet. We obtained 0.79 and 0.78 onSVR for the nDCG@5 and RankNet for thenDCG@10, respectively. We focus on thesubject word in each sentence. The methodwith subject information outperformed thatwithout subject information; 0.79 with subjectinformation vs. 0.64 without subject information on the SVR for the nDCG@5.

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85083277805&origin=inward

Discussion map with an assistant function for decision-making: A tool for supporting consensus-building 査読有り国際誌

Kirikihira R., Shimada K.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11000 LNCS 3 - 18 2018年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

In this paper, we propose a tool for supporting consensus-building in conversations with multiple participants. We call it “Discussion Map with Assistant (DMA)”. It consists of nodes and links. We classify the nodes into two types; alternatives and criteria. Alternatives represent what the participants are choosing between. Criteria are used to judge the alternatives. Each criterion contains an importance value. Each link between nodes also contains an importance value. The system estimates a ranking list of alternatives among participants from each map. We introduce a forgetting function to the model. The system also supports the decision-making process by using discussion maps from participants. It generates sentences and charts that describe the current state of the discussion. We evaluate the effectiveness of the discussion map system with DMA in a decision-making task experimentally.

DOI： 10.1007/978-3-319-98743-9_1

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85053106635&origin=inward

Analysis of facilitators’ behaviors in multi-party conversations for constructing a digital facilitator system 査読有り国際誌

Shiota T., Yamamura T., Shimada K.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11000 LNCS 145 - 158 2018年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

In this paper, we analyze characteristics of facilitators from multi-party conversations. The goal of our study is to construct a digital facilitator system that supports consensus-building and management of conversation for high-quality discussions. Therefore, we need facilitator’s knowledge, behavior, and patterns to realize a good digital facilitator. As the 1st step for the purpose, we focus on a macro viewpoint of facilitators’ behaviors on conversation corpora. First, we generate a model based on a decision tree that classifies each participant into a facilitator or a non-facilitator, from conversation corpora. The classification accuracies by the decision trees were 0.642 and 0.737 for two corpora, respectively. The main purpose of the decision tree generation is to extract patterns from imaginable characteristics, namely features for the classifier. Therefore, next, we discuss behaviors of facilitators by analyzing the decision tree manually. In the analysis, we focus on two types of corpora; one is that each participant has a role, such as a project manager, and another is that each participant has no role in the conversation. We investigate the influence of the difference of the setting through the analysis. From the manual analysis, we obtained some common behaviors and some different behaviors about facilitators from two corpora.

DOI： 10.1007/978-3-319-98743-9_12

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85053169078&origin=inward

　前のページ - 次のページ　

<嶋田 和孝>

論文 - 嶋田 和孝

<嶋田　和孝>

論文 - 嶋田　和孝