Knowledge Management and Acquisition for Intelligent Systems: 16th Pacific Rim Knowledge Acquisition Workshop, PKAW 2019, Cuvu, Fiji, August 26–27, 2019, Proceedings [1st ed. 2019] 978-3-030-30638-0, 978-3-030-30639-7

Table of contents :
Front Matter ....Pages i-x
Estimating Difficulty Score of Visual Search in Images for Semi-supervised Object Detection (Dongliang Ma, Haipeng Zhang, Hao Wu, Tao Zhang, Jun Sun)....Pages 1-9
Improving Named Entity Recognition with Commonsense Knowledge Pre-training (Ghaith Dekhili, Ngoc Tan Le, Fatiha Sadat)....Pages 10-20
Neurofeedback and AI for Analyzing Child Temperament and Attention Levels (Maria R. Lee, Anna Yu-Ju Yen, Lun Chang)....Pages 21-31
Finding Diachronic Objects of Drifting Descriptions by Similar Mentions (Katsuaki Tanaka, Koichi Hori)....Pages 32-43
A Max-Min Conflict Algorithm for the Stable Marriage Problem (Hoang Huu Viet, Nguyen Thi Uyen, Pham Tra My, Son Thanh Cao, Le Hong Trang)....Pages 44-53
Empirical Evaluation of Deep Learning-Based Travel Time Prediction (Mengyan Wang, Weihua Li, Yan Kong, Quan Bai)....Pages 54-65
Marine Vertebrate Predator Detection and Recognition in Underwater Videos by Region Convolutional Neural Network (Mira Park, Wenli Yang, Zehong Cao, Byeong Kang, Damian Connor, Mary-Anne Lea)....Pages 66-80
Constructing Dataset Based on Concept Hierarchy for Evaluating Word Vectors Learned from Multisense Words (Tomoaki Yamazaki, Tetsuya Toyota, Kouzou Ohara)....Pages 81-96
Adaptive Database’s Performance Tuning Based on Reinforcement Learning (Chee Keong Wee, Richi Nayak)....Pages 97-114
Prior-Knowledge-Embedded LDA with Word2vec – for Detecting Specific Topics in Documents (Hiroshi Uehara, Akihiro Ito, Yutaka Saito, Kenichi Yoshida)....Pages 115-126
Comparative Analysis of Intelligent Personal Agent Performance (David Herbert, Byeong Kang)....Pages 127-141
Toxicity Prediction by Multimodal Deep Learning (Abdul Karim, Jaspreet Singh, Avinash Mishra, Abdollah Dehzangi, M. A. Hakim Newton, Abdul Sattar)....Pages 142-152
Context-Aware Influence Diffusion in Online Social Networks (Yuxuan Hu, Quan Bai, Weihua Li)....Pages 153-162
Network Embedding via Link Strength Adjusted Random Walk (Chenliang Li, Donghai Guan, Weiwei Yuan)....Pages 163-172
Study on Influencers of Cryptocurrency Follow-Network on GitHub (Naoki Kobayakawa, Kenichi Yoshida)....Pages 173-183
A Cross-Domain Theory of Mental Models (Sara Todorovikj, Paulina Friemann, Marco Ragni)....Pages 184-194
Back Matter ....Pages 195-195

Citation preview

LNAI 11669

Kouzou Ohara Quan Bai (Eds.)

Knowledge Management and Acquisition for Intelligent Systems 16th Pacific Rim Knowledge Acquisition Workshop, PKAW 2019 Cuvu, Fiji, August 26–27, 2019 Proceedings

123

Lecture Notes in Artificial Intelligence Subseries of Lecture Notes in Computer Science

Series Editors Randy Goebel University of Alberta, Edmonton, Canada Yuzuru Tanaka Hokkaido University, Sapporo, Japan Wolfgang Wahlster DFKI and Saarland University, Saarbrücken, Germany

Founding Editor Jörg Siekmann DFKI and Saarland University, Saarbrücken, Germany

11669

More information about this series at http://www.springer.com/series/1244

Kouzou Ohara Quan Bai (Eds.) •

Knowledge Management and Acquisition for Intelligent Systems 16th Pacific Rim Knowledge Acquisition Workshop, PKAW 2019 Cuvu, Fiji, August 26–27, 2019 Proceedings

123

Editors Kouzou Ohara Aoyama Gakuin University Tokyo, Japan

Quan Bai University of Tasmania Tasmania, Australia

ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Artificial Intelligence ISBN 978-3-030-30638-0 ISBN 978-3-030-30639-7 (eBook) https://doi.org/10.1007/978-3-030-30639-7 LNCS Sublibrary: SL7 – Artificial Intelligence © Springer Nature Switzerland AG 2019 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

This volume contains the papers presented at the 2019 Pacific Rim Knowledge Acquisition Workshop (PKAW 2019) held in conjunction with the 16th Pacific Rim International Conference on Artificial Intelligence (PRICAI 2019), during August 26–27, 2019, in Fiji. Since the 1990s, PKAW has provided researchers with opportunities to present ideas and have intensive discussions on their work related to knowledge acquisition, which is one of core fields of Artificial Intelligence (AI). The scope of PKAW is not limited to traditional knowledge acquisition approaches such as human (expert) centric ones, but also covers diverse areas closely related to knowledge acquisition such as knowledge engineering, knowledge management, machine learning, data mining, etc. We need to choose appropriate techniques for knowledge acquisition, depending on its type and the task addressed. In fact, the scope changes over time so that it can cover novel, newly emerged techniques and application areas in which knowledge acquisition plays an important role. Especially, now, we live in the era of the third wave of AI, in which the availability of high-performance computing and massive electronic data generated from various sensors and texts on the Web make it possible to devise novel data-driven methodologies such as Deep Learning and its variants. These advanced technologies could help us acquire tacit knowledge that has been difficult to learn by human-centric approaches, while they also remind us of the importance of the understandability of knowledge, leading to a new field of Explainable AI. Within this context, we invited submissions in the above broad fields and finally selected 9 regular papers and 7 short papers from 38 submitted papers. All papers were peer-reviewed by three reviewers. These papers demonstrate advanced research work from the practical viewpoint and make contributions in technical and theoretical aspects to the fields of intelligent systems/agents, natural language processing, and applications of machine learning techniques including Deep Learning to real world problems. The workshop co-chairs would like to thank all the people who supported PKAW 2019, including the PKAW Program Committee members and sub-reviewers who spent their precious time reviewing papers, the PRICAI Organizing Committee who appropriately dealt with our requests and all of the administrative and local matters. Thanks to EasyChair for providing the online platform to efficiently handle submissions and to Springer for publishing the proceedings in the Lecture Notes in Artificial Intelligence (LNAI) series. Of course, we would like to give a special thanks to all authors who submitted papers, all presenters, and all attendees. August 2019

Kouzou Ohara Quan Bai

Organization

Organizing Committee Honorary Chairs Paul Compton Hiroshi Motoda

The University of New South Wales, Australia Osaka University, Japan

Workshop Co-chairs Kouzou Ohara Quan Bai

Aoyama Gakuin University, Japan University of Tasmania, Australia

Publicity Chair Soyeon Caren Han

University of Sydney, Australia

Advisory Committee Maria R. Lee Kenichi Yoshida Byeong-Ho Kang Deborah Richards

Shih Chien University, Taiwan University of Tsukuba, Japan University of Tasmania, Australia Macquarie University, Australia

Program Committee Xiongcai Cai Zehong Cao Tsung-Teng Chen Akihiro Inokuchi Toshihiro Kamishima Alfred Krzywicki Weihua Li Toshiro Minami Tsuyoshi Murata Hayato Ohwada Tomonobu Ozaki Hye-Young Paik Mira Park Ulrich Reimer Kazumi Saito Derek Sleeman

The University of New South Wales, Australia University of Tasmania, Australia NTPU, Taiwan Kwansei Gakuin University, Japan National Institute of Advanced Industrial Science and Technology (AIST), Japan The University of New South Wales, Australia Auckland University of Technology, New Zealand Kyushu Institute of Information Sciences, Kyushu University Library, Japan Tokyo Institute of Technology, Japan Tokyo University of Science, Japan Nihon University, Japan The University of New South Wales, Australia University of Tasmania, Australia University of Applied Sciences St. Gallen, Switzerland University of Shizuoka, Japan University of Aberdeen, UK

viii

Organization

Xing Su Vojtěch Svátek Hiroshi Uehara Shuxiang Xu Takahira Yamaguchi Tetsuya Yoshida

Additional Reviewer Hahn, Heiko

Beijing University of Technology, China University of Economics Prague, Czech Republic Akita Prefectural University, Japan University of Tasmania, Australia Keio University, Japan Nara Women’ University, Japan

Contents

Estimating Difficulty Score of Visual Search in Images for Semi-supervised Object Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dongliang Ma, Haipeng Zhang, Hao Wu, Tao Zhang, and Jun Sun

1

Improving Named Entity Recognition with Commonsense Knowledge Pre-training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ghaith Dekhili, Ngoc Tan Le, and Fatiha Sadat

10

Neurofeedback and AI for Analyzing Child Temperament and Attention Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Maria R. Lee, Anna Yu-Ju Yen, and Lun Chang

21

Finding Diachronic Objects of Drifting Descriptions by Similar Mentions . . . Katsuaki Tanaka and Koichi Hori

32

A Max-Min Conflict Algorithm for the Stable Marriage Problem . . . . . . . . . Hoang Huu Viet, Nguyen Thi Uyen, Pham Tra My, Son Thanh Cao, and Le Hong Trang

44

Empirical Evaluation of Deep Learning-Based Travel Time Prediction. . . . . . Mengyan Wang, Weihua Li, Yan Kong, and Quan Bai

54

Marine Vertebrate Predator Detection and Recognition in Underwater Videos by Region Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . Mira Park, Wenli Yang, Zehong Cao, Byeong Kang, Damian Connor, and Mary-Anne Lea

66

Constructing Dataset Based on Concept Hierarchy for Evaluating Word Vectors Learned from Multisense Words . . . . . . . . . . . . . . . . . . . . . . . . . . Tomoaki Yamazaki, Tetsuya Toyota, and Kouzou Ohara

81

Adaptive Database’s Performance Tuning Based on Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chee Keong Wee and Richi Nayak

97

Prior-Knowledge-Embedded LDA with Word2vec – for Detecting Specific Topics in Documents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hiroshi Uehara, Akihiro Ito, Yutaka Saito, and Kenichi Yoshida

115

Comparative Analysis of Intelligent Personal Agent Performance . . . . . . . . . David Herbert and Byeong Kang

127

x

Contents

Toxicity Prediction by Multimodal Deep Learning . . . . . . . . . . . . . . . . . . . Abdul Karim, Jaspreet Singh, Avinash Mishra, Abdollah Dehzangi, M. A. Hakim Newton, and Abdul Sattar

142

Context-Aware Influence Diffusion in Online Social Networks . . . . . . . . . . . Yuxuan Hu, Quan Bai, and Weihua Li

153

Network Embedding via Link Strength Adjusted Random Walk . . . . . . . . . . Chenliang Li, Donghai Guan, and Weiwei Yuan

163

Study on Influencers of Cryptocurrency Follow-Network on GitHub . . . . . . . Naoki Kobayakawa and Kenichi Yoshida

173

A Cross-Domain Theory of Mental Models . . . . . . . . . . . . . . . . . . . . . . . . Sara Todorovikj, Paulina Friemann, and Marco Ragni

184

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

195

Estimating Difficulty Score of Visual Search in Images for Semi-supervised Object Detection Dongliang Ma, Haipeng Zhang, Hao Wu, Tao Zhang, and Jun Sun(B) JiangSu Provincial Engineering Laboratory of Pattern Recognition and Computational Intelligence, JiangNan University, Wuxi, China [email protected], [email protected], [email protected], {taozhang,junsun}@jiangnan.edu.cn

Abstract. Humans can intuitively understand the content of images, and often reach a consensus that some images are more difficult to visual search tasks than others. However, this is quite challenging for computers as it is a subjective task which may be influenced by human emotional factors. Instead of focusing on how the models make reactions on datasets, our method has a capability of assigning scores to samples respectively within a dataset that estimating the difficulty of visual search tasks. Our model shows better performance for predicting visual search difficulty scores of samples produced by human annotators in PASCAL VOC2012. Eventually, we demostrate with experiments that our method has an ability of selecting suitable samples to improve the performance of detectors in a semi-supervised task.

Keywords: Visual search

1

· Semi-supervised object detection

Introduction

Since its development two decades ago, Convolutional Neural Networks have promoted the development of computer vision. Various network architectures [4,7,11] all prove the power of CNNs. On the other hand, cognitive research [1,13,16] shows that for tasks that search for patterns in images, response time of users make effects on visual search difficulty, and visual search difficulty may vary from image to image. Salient objects could be quickly found in images, while others are more difficult, requiring humans to perform intensive visual processing as shown in Fig. 1. Regrettably, CNNs still can’t think and analyze problems in human way. In previous work [5,9,14], they addressed the issue of estimating image difficulty defined as the response time of human for visual search tasks. The task may be affected by the following factors, i.e. the background, complexity of scene, amount of objects, and possible occlusion. In this work, we have established a new end-to-end neural network regression model to predict the difficulty scores c Springer Nature Switzerland AG 2019  K. Ohara and Q. Bai (Eds.): PKAW 2019, LNAI 11669, pp. 1–9, 2019. https://doi.org/10.1007/978-3-030-30639-7_1

2

D. Ma et al.

Fig. 1. Some figures captured by PASCAL VOC2012 are shown in visual search tasks. The object to be identified in the first line is the aircraft, the second birds, the third people. For the same object, we can easily find it in the left image, which is difficulty to find in the image on the right. The difficulty of the picture increases from left to right in our mind. Obviously, the time it takes to find the specified object in the above images are not the same.

of images, so that the model can automatically predict the human’s assessment of the difficulty of images. Moreover, almost all of the existed works investigate only the models and ignore the relationship between models and samples, estimating the difficulty of images can be used as a criterion for distinguishing images. It has been employed in some previous works for weakly supervised object detection [10, ?], semisupervised object classification [5], and potential application in object detection [12,15]. In order to verify the strength of our work, we use the difficulty score of image represents both confidence and interpretability that we can select some particular samples to improve model detection accuracy. In terms of the difficulty estimation for visual search in images, this paper has the following two contributions. Firstly, Kendall RankLoss is designed for optimizing the Kendall’s τ correlation coefficient [3] which evaluates the neural network to estimate the difficulty scores of images. Secondly, we utilize our score of difficulty as a indication to label the unlabeled data during training in semisupervised object detection. The rest of the paper is organized as follows. Section 2 introduces related works and Sect. 3 presents our proposed work. In Sect. 4, the experimental results are reported and analyzed. Lastly, some concluding remarks are given in Sect. 5.

Estimating Difficulty Score of Visual Search

2

3

Related Work

At present, there are few researches conducted on image difficulty [5,9,14]. Russakovsky et al. [9] measured difficulty by the size and the number of boundingboxes (even at test time). Grauman and Vijayanarasimhan [14] attempted to evaluate the difficulty of an image based on the time it takes for humans to segment it. However the task of image segmentation [14] is fundamentally different from our visual search task. For example, a truncated object can be easily segmented but difficult to be found and identified. Radu et al. [5] used the specially established PASCAL VOC2012 dataset close to human visual level to estimate the difficulty of images. The process of dataset labelling was performed on a crowd-sourcing platform named CrowdFlower, after 736 trusted annotators observe the information in the image, and the time required to answer the question was used as a measure of the image difficulty and convert it into image difficulty score. They designed a series of related processing methods including clearing outliers to ensure the validity of the data. Based on the usual visual search tasks, they proposed an explanation of the difficulty of the image close to the human cognitive level. Since the background of each image in the PASCAL VOC2012 dataset is different, the densities, numbers, sizes and appearances of the objects are different, which can meet the demand of experimental validation. They built a regression model based on deep features learned by convolutional neural network. Importantly, they used MSE (Mean Squared Error) and Kendall’s τ correlation coefficient to evaluate the difficulty generalization performance of the model. In our work, we designed the Kendall RankLoss for optimizing the Kendall’s τ correlation coefficient [3]. In addition to what is different that they regarded the neural network as a feature extractor to get information, we employ the same CNN as an end-to-end regression model directly. We validate our model with the same dataset and evaluation criteria as they do.

3 3.1

Our Approach Kendall RankLoss

When the model is trained, if the batch size of the neural network is large enough, we can back-propagate the information to the network with the difference between the rank information of the predicted difficulty scores and the ground-truth rank information in each batch as the error, and resultantly the entire network can continuously reduce the error. By the definition of Kendall’s τ correlation coefficient, the range of its value is between −1 and 1. When the rank information of all samples is exactly the same as the rank information of the real values, Kendall’s τ correlation coefficient takes a value of 1, and vice versa. Supposedly, there are n samples for each mini-batch {xi , yti }N i=1 , where xi represents the i-th image in the batch, yti denotes the true label and ypi is known as the corresponding predicted difficulty score by model, and the sgn function is denoted by f . Therefore, we can define Kendall’s τ correlation coefficient on the batch as follows

4

D. Ma et al.



 τ=

i∈N

(j