Trends in Artificial Intelligence

Full download test bank at ebook textbookfull.
com
Trends in Artificial Intelligence

Theory and Applications. Artificial
Intelligence Practices: 33rd
International Conference on
Industrial, Engineering and Other
CLICK LINK TO DOWLOAD
Applications of Applied Intelligent
https://textbookfull.com/product/trends-in-
Systems, IEA/AIE 2020, Kitakyushu,
artificial-intelligence-theory-and-
Japan,applications-artificial-intelligence-
September 22-25, Hamido
practices-33rd-international-conference-on-
industrial-engineering-and-other-
applications-of-applied-intelligent-sys/
textbookfull
Hamido Fujita
Philippe Fournier-Viger
Moonis Ali
Jun Sasaki (Eds.)
Trends in Artificial Intelligence

LNAI 12144
Theory and Applications

Artificial Intelligence Practices
33rd International Conference
on Industrial, Engineering and Other Applications
of Applied Intelligent Systems, IEA/AIE 2020
Kitakyushu, Japan, September 22–25, 2020, Proceedings
123
Lecture Notes in Artificial Intelligence 12144
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
More information about this series at http://www.springer.com/series/1244
Hamido Fujita Philippe Fournier-Viger
• •
Moonis Ali Jun Sasaki (Eds.)

•
Trends in Artificial Intelligence

Theory and Applications
Artificial Intelligence Practices
33rd International Conference
on Industrial, Engineering and Other Applications
of Applied Intelligent Systems, IEA/AIE 2020
Kitakyushu, Japan, September 22–25, 2020
Proceedings
123
Editors
Hamido Fujita Philippe Fournier-Viger
Iwate Prefectural University Harbin Institute of Technology (Shenzhen)
Takizawa, Japan Shenzhen, China
Moonis Ali Jun Sasaki
Texas State University Iwate Prefectural University
San Marcos, TX, USA Takizawa, Japan
ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Artificial Intelligence
ISBN 978-3-030-55788-1 ISBN 978-3-030-55789-8 (eBook)
https://doi.org/10.1007/978-3-030-55789-8
LNCS Sublibrary: SL7 – Artificial Intelligence
© Springer Nature Switzerland AG 2020

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
In recent decades, society has entered a digital era where computers have become
ubiquitous in all aspects of life, including education, governance, science, healthcare,
and industry. Computers have become smaller, faster and the cost of data storage and
communication have greatly decreased. As a result, more and more data is being
collected and stored in databases. Besides, novel and improved computing architectures
have been designed for efficient large-scale data processing such as big data frame-
works, FPGAs, and GPUs. Thanks to these advancements and recent breakthroughs in
artificial intelligence, researchers and practitioners have developed more complex and
effective artificial intelligence-based systems. This has led to a greater interest in
artificial intelligence to solve real-world complex problems, and the proposal of many
innovative applications.
This volume contains the proceedings of the 33rd edition of the International
Conference on Industrial, Engineering, and other Applications of Applied Intelligent
Systems (IEA AIE 2020), which was held during September 22–25, 2020, in
Kitakyushu, Japan. IEA AIE is an annual event that emphasizes applications of applied
intelligent systems to solve real-life problems in all areas including engineering, sci-
ence, industry, automation and robotics, business and finance, medicine and biome-
dicine, bioinformatics, cyberspace, and human-machine interactions. This year, 119
submissions were received. Each paper was evaluated using a double-blind peer review
by at least three reviewers from an international Program Committee consisting of 82
members from 36 countries. Based on the evaluation, a total of 62 papers were selected
as full papers and 17 as short papers, which are presented in this book. We would like
to thank all the reviewers for the time spent to write detailed and constructive com-
ments to authors, and to these latter for the proposal of many high-quality papers.
In the program of IEA AIE 2020, two special sessions were organized named
Collective Intelligence in Social Media (CISM 2020) and Intelligent Knowledge
Engineering in Decision Making Systems (IKEDS 2020). Moreover, three keynote
talks were given by distinguished researchers, one by Prof. Tao Wu from Shanghai Jiao
Tong University School of Medicine (China), one by Enrique Herrera Viedma from the
University of Granada (Spain), and another by Ee-Peng Lim from Singapore Man-
agement University (Singapore). Lastly, we would like to thank everyone who con-
tributed to the success of this year’s edition of IEA AIE that is authors, Program
Committee members, reviewers, keynote speakers, and organizers.
September 2020 Hamido Fujita

Philippe Fournier-Viger
Moonis Ali
Jun Sasaki
Organization
General Chair
Hamido Fujita Iwate Prefectural University, Japan
General Co-chairs
Moonis Ali Texas State University, USA
Franz Wotawa TU Graz, Austria
Organizing Chair
Jun Sasaki Iwate Prefectural University, Japan
Program Chairs
Philippe Fournier-Viger Harbin Institute of Technology (Shenzhen), China
Hideyuki Takagi Kyushu University, Japan
Special Session Chairs

Yinglin Wang Shanghai University of Finance and Economic, China
Ali Selamat Universiti Teknologi Malaysia, Malaysia
Prima O.D.A. Iwate Prefectural University, Japan
Special Session Organizers

Jerry Chun-Wei Lin Western Norway University of Applied Sciences,
Norway
Rage Uday Kiran University of Aizu, Japan
Ngoc-Thanh Nguyen Wroclaw University of Science and Technology,
Poland
Van Du Nguyen Nong Lam University, Vietnam
Publicity Chair
Toshitaka Hayashi Iwate Prefectural University, Japan
viii Organization
Program Committee
Rui Abreu University of Lisbon, Portugal
Otmane Ait Mohamed Corcordia University, Canada
Hadjali Allel ENSMA, France
Xiangdong An The University of Tennessee, USA
Artur Andrzejak Heidelberg University, Germany
Farshad Badie Aalborg University, Denmark
Ladjel Bellatreche ENSMA, France
Fevzi Belli Paderborn University, Germany
Adel Bouhoula University of Carthage, Tunisia
Ivan Bratko University of Ljubljana, Slovenia
João Paulo Carvalho University of Lisbon, Portugal
Chun-Hao Chen National Taipei University of Technology, Taiwan
Shyi-Ming Chen National Taiwan University of Science
and Technology, Taiwan
Flávio Soares Corrêa da University of São Paulo, Brazil
Silva
Giorgos Dounias University of the Aegean, Greece
Alexander Ferrein Aachen University of Applied Science, Germany
Hamido Fujita Iwate Prefectural University, Japan
Vicente García Díaz University of Oviedo, Spain
Alban Grastien The Australian National University, Australia
Maciej Grzenda Warsaw University of Technology, Poland
Jun Hakura Iwate Prefectural University, Japan
Tim Hendtlass School of Biophysical Sciences and Electrical
Engineering, Australia
Dinh Tuyen Hoang Yeungnam University, South Korea
Tzung-Pei Hong National University of Kaohsiung, Taiwan
Wen-Juan Hou National Central University, Taiwan
Ko-Wei Huang National Kaohsiung University of Science
and Technology, Taiwan
Quoc Bao Huynh Ho Chi Minh City University of Technology, Vietnam
Said Jabbour University of Artois, France
He Jiang Dalian University of Technology, China
Rage Uday Kiran University of Aizu, Japan
Yun Sing Koh The University of Auckland, New Zealand
Adrianna Kozierkiewicz Wroclaw University of Science and Technology,
Poland
Dariusz Krol Wroclaw University of Science and Technology,
Poland
Philippe Leray University of Nantes, France
Mark Levin Russian Academy of Sciences, Russia
Jerry Chun-Wei Lin Western Norway University of Applied Sciences,
Norway
Organization ix
Yu-Chen Lin Feng Chia University, Taiwan

Jose Maria-Luna University of Cordoba, Spain
Wolfgang Mayer University of South Australia, Australia
Joao Mendes-Moreira University of Porto, Portugal
Engelbert Mephu Nguifo Université Clermont Auvergne, France
Mercedes Merayo Universidad Complutense de Madrid, Spain
Abidalrahman Moh’D Eastern Illinois University, USA
Anirban Mondal Ashoka University, India
Saqib Nawaz Harbin Institute of Technology (Shenzhen), China
Roger Nkambou Université du Québec à Montréal, Canada
Ngoc-Thanh Nguyen Wroclaw University of Science and Technology,
Poland
Quang Vu Nguyen Vietnam-Korea Friendship Information Technology
College, Vietnam
Van Du Nguyen Nong Lam University, Vietnam
Ayahiko Niimi Future University Hakodate, Japan
Xinzheng Niu University of Electronic Science and Technology
of China, China
Farid Nouioua Aix-Marseille Université, France
Mourad Nouioua Harbin Institute of Technology (Shenzhen), China
Barbara Pes University of Cagliari, Italy
Marcin Pietranik Wroclaw University of Science and Technology,
Poland
Ingo Pill TU Graz, Austria
Matin Pirouz California State University, USA
Krishna P. Reddy International Institute of Information Technology,
Hyderabad, India
Gregorio Sainz-Palmero University of Valladolid, Spain
Eugene Santos Jr. Dartmouth College, USA
Jun Sasaki Iwate Prefectural University, Japan
Ali Selamat Universiti Teknologi Malaysia, Malaysia
Nazha Selmaoui-Folcher University of New Caledonia, New Caledonia
Sabrina Senatore University of Salerno, Italy
Neal Snooke Aberystwyth University, UK
Gerald Steinbauer TU Graz, Austria
Ahmed Tawfik Microsoft Research, USA
Trong Hieu Tran Hanoi University of Engineering and Technology,
Vietnam
Van Cuong Tran Quang Binh University, Vietnam
Chun-Wei Tsai National Sun Yat-sen University, Taiwan
Alexander Vazhenin University of Aizu, Japan
Bay Vo HCM City University of Technology, Vietnam
Toby Walsh NICTA, Australia
Yutaka Watanobe University of Aizu, Japan
Tomasz Wiktorski University of Stavanger, Norway
Cheng Wei Wu National Ilan University, Taiwan
x Organization
Franz Wotawa TU Graz, Austria

Jimmy Ming-Tai Wu Shandong University of Science and Technology,
China
Mu-En Wu National Taipei University of Technology, Taiwan
Unil Yun Sejong University, South Korea
Wei Zhang Adobe Systems, USA
Contents
Natural Language Processing
Question Generation Through Transfer Learning . . . . . . . . . . . . . . . . . . . . . 3

Yin-Hsiang Liao and Jia-Ling Koh
KIDER: Knowledge-Infused Document Embedding Representation

for Text Categorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Yu-Ting Chen, Zheng-Wen Lin, Yung-Chun Chang, and Wen-Lian Hsu
Discriminative Features Fusion with BERT for Social Sentiment Analysis . . . 30

Duy-Duc Le Nguyen, Yen-Chun Huang, and Yung-Chun Chang
Text Sentiment Transfer Methods by Using Sentence Keywords . . . . . . . . . . 36

Shengwei Hu, Bicheng Li, Kongjie Lin, Rui Wang, and Kai Liu
Robotics and Drones
Path Planning of Mobile Robot Group Based on Neural Networks . . . . . . . . 51

Mikhail Medvedev and Viacheslav Pshikhopov
Push Recovery and Active Balancing for Inexpensive Humanoid Robots

Using RL and DRL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Amirhossein Hosseinmemar, John Anderson, Jacky Baltes,
Meng Cheng Lau, and Ziang Wang
Optimal Control Problem of a Differential Drive Robot . . . . . . . . . . . . . . . . 75

Luis F. Recalde, Bryan S. Guevara, Giovanny Cuzco,
and Víctor H. Andaluz
Optimal Trajectory Tracking Control for a UAV Based on Linearized

Dynamic Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Christian P. Carvajal, Víctor H. Andaluz, Flavio Roberti,
and Ricardo Carelli
Non-linear Control of Aerial Manipulator Robots Based

on Numerical Methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
David F. Grijalva, Jaime A. Alegría, Víctor H. Andaluz,
and Cesar Naranjo
Non-linear 3D Visual Control for an Unmanned Aerial Vehicle . . . . . . . . . . 108

Daniel D. Guevara and Víctor H. Andaluz
xii Contents
Construction and Control Aerial Manipulator Robot . . . . . . . . . . . . . . . . . . 116

Steeven J. Loor, Alan R. Bejarano, Franklin M. Silva,
and Víctor H. Andaluz
Knowledge Based Systems
ConMerge – Arbitration of Constraint-Based Knowledge Bases . . . . . . . . . . 127

Mathias Uta and Alexander Felfernig
A Systematic Model to Model Transformation for Knowledge-Based

Planning Generation Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Liwen Zhang, Franck Fontanili, Elyes Lamine, Christophe Bortolaso,
Mustapha Derras, and Hervé Pingaud
Innovative Applications of Intelligent Systems
Mathematical Expression Retrieval in PDFs from the Web Using

Mathematical Term Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Kuniko Yamada and Harumi Murakami
Automatic Identification of Account Sharing for Video Streaming Services. . . 162

Wei Zhang and Chris Challis
A Model for Predicting Terrorist Network Lethality and Cohesiveness. . . . . . 174

Botambu Collins, Dinh Tuyen Hoang, and Dosam Hwang
S2RSCS: An Efficient Scientific Submission Recommendation System

for Computer Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Son T. Huynh, Phong T. Huynh, Dac H. Nguyen, Dinh V. Cuong,
and Binh T. Nguyen
Improved Grey Model by Dragonfly Algorithm for Chinese Tourism

Demand Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Jinran Wu and Zhe Ding
Variable Transformation to a 2 2 Domain Space for Edge

Matching Puzzles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
Thomas Aspinall, Adrian Gepp, Geoff Harris,
and Bruce James Vanstone
Industrial Applications
Using Deep Learning Techniques to Detect Rice Diseases from Images

of Rice Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Kantip Kiratiratanapruk, Pitchayagan Temniranrat,
Apichon Kitvimonrat, Wasin Sinthupinyo, and Sujin Patarapuwadol
Contents xiii
Machine Learning for Water Supply Supervision . . . . . . . . . . . . . . . . . . . . 238

Thomas Schranz, Gerald Schweiger, Siegfried Pabst, and Franz Wotawa
An Enhanced Whale Optimization Algorithm for the Two-Dimensional

Irregular Strip Packing Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
Qiang Liu, Zehui Huang, Hao Zhang, and Lijun Wei
A Heuristic Approach to the Three Dimensional Strip Packing Problem

Considering Practical Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
Qiang Liu, Dehao Lin, Hao Zhang, and Lijun Wei
A Heuristic for the Two-Dimensional Irregular Bin Packing Problem

with Limited Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Qiang Liu, Jiawei Zeng, Hao Zhang, and Lijun Wei
Faster R-CNN Based Fault Detection in Industrial Images . . . . . . . . . . . . . . 280

Faisal Saeed, Anand Paul, and Seungmin Rho
Estimation of Cable Lines Insulating Materials Resource Using Multistage

Neural Network Forecasting Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
Nikolay K. Poluyanovich, Mikhail Yu Medvedev, Marina N. Dubyago,
Nikolay V. Azarov, and Alexander V. Ogrenichev
Networking Applications
User Grouping and Power Allocation in NOMA Systems: A Reinforcement

Learning-Based Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
Rebekka Olsson Omslandseter, Lei Jiao, Yuanwei Liu,
and B. John Oommen
Deep Learning for QoS-Aware Resource Allocation in Cognitive

Radio Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
Jerzy Martyna
Social Network Analysis
Self-understanding Support Tool Using Twitter Sentiment Analysis. . . . . . . . 327

Harumi Murakami, Naoya Ejima, and Naoto Kumagai
Integrating Crowdsourcing and Active Learning for Classification

of Work-Life Events from Tweets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Yunpeng Zhao, Mattia Prosperi, Tianchen Lyu, Yi Guo, Le Zhou,
and Jiang Bian
Many-to-One Stable Matching for Prediction in Social Networks . . . . . . . . . 345

Ke Dong, Zengchang Qin, and Tao Wan
xiv Contents
A Framework for Detecting User’s Psychological Tendencies on Twitter

Based on Tweets Sentiment Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
Huyen Trang Phan, Van Cuong Tran, Ngoc Thanh Nguyen,
and Dosam Hwang
Automatic Fake News Detection by Exploiting User’s Assessments

on Social Networks: A Case Study of Twitter. . . . . . . . . . . . . . . . . . . . . . . 373
Van Cuong Tran, Van Du Nguyen, and Ngoc Thanh Nguyen
Financial Applications and Blockchain
Deep Reinforcement Learning for Foreign Exchange Trading . . . . . . . . . . . . 387

Yun-Cheng Tsai, Chun-Chieh Wang, Fu-Min Szu, and Kuan-Jen Wang
Human-Centred Automated Reasoning for Regulatory Reporting

via Knowledge-Driven Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
Dilhan J. Thilakarathne, Newres Al Haider, and Joost Bosman
Security of Blockchain Distributed Ledger Consensus Mechanism

in Context of the Sybil Attack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Michal Kedziora, Patryk Kozlowski, and Piotr Jozwiak
Reinforcement Learning Based Real-Time Pricing in Open Cloud Markets. . . 419

Pankaj Mishra, Ahmed Moustafa, and Takayuki Ito
Medical and Health-Related Applications
A New Integer Linear Programming Formulation to the Inverse QSAR/

QSPR for Acyclic Chemical Compounds Using Skeleton Trees . . . . . . . . . . 433
Fan Zhang, Jianshen Zhu, Rachaya Chiewvanichakorn,
Aleksandar Shurbevski, Hiroshi Nagamochi, and Tatsuya Akutsu
Computing a Weighted Jaccard Index of Electronic Medical Record

for Disease Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
Chia-Hui Huang, Yun-Te Liao, David Taniar, and Tun-Wen Pai
The Differential Feature Detection and the Clustering Analysis

to Breast Cancers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Juanying Xie, Zhaozhong Wu, Qin Xia, Lijuan Ding, and Hamido Fujita
Left Ventricle Segmentation Using Scale-Independent Multi-Gate UNET

in MRI Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
Mina Saber, Dina Abdelrauof, and Mustafa Elattar
Clustering-Based Data Reduction Approach to Speed up SVM

in Classification and Regression Tasks. . . . . . . . . . . . . . . . . . . . . . . . . . . . 478
Adamo Santana, Souta Inoue, Kenya Murakami, Tatsuya Iizaka,
and Tetsuro Matsui
Contents xv
AI for Health – Knowledge-Based Generation of Tailor-Made

Exercise Plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
Florian Grigoleit, Peter Struss, and Florian Kreuzpointner
Anomaly Detection and Automated Diagnosis
A Multi-phase Iterative Approach for Anomaly Detection and Its

Agnostic Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
Kévin Ducharlet, Louise Travé-Massuyès, Marie-Véronique Le Lann,
and Youssef Miloudi
On the Use of Answer Set Programming for Model-Based Diagnosis . . . . . . 518

Franz Wotawa
Decision-Support and Agent-Based Systems
Consensus-Based Protocol for Distributed Exploration and Mapping . . . . . . . 533

Zilong Jiao and Jae Oh
A Real-Time Actor-Critic Architecture for Continuous Control . . . . . . . . . . . 545

Zilong Jiao and Jae Oh
Action-Based Programming with YAGI - An Update on Usability

and Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
Thomas Eckstein and Gerald Steinbauer
A New Approach to Determine 2-Optimality Consensus for Collectives. . . . . 570

Dai Tho Dang, Zygmunt Mazur, and Dosam Hwang
A Decision Support System to Provide Criminal Pattern Based Suggestions

to Travelers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
Khin Nandar Win, Jianguo Chen, Mingxing Duan, Guoqing Xiao,
Kenli Li, Philippe Fournier-Viger, and Keqin Li
Model-Based Decision Support Systems - Conceptualization

and General Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
Peter Struss
Multimedia Applications
Driver License Field Detection Using Real-Time Deep Networks . . . . . . . . . 603

Chun-Ming Tsai, Jun-Wei Hsieh, Ming-Ching Chang, and Yu-Chen Lin
Calibration of a Microphone Array Based on a Probabilistic Model

of Microphone Positions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
Katsuhiro Dan, Katsutoshi Itoyama, Kenji Nishida,
and Kazuhiro Nakadai
xvi Contents
How to Handle Head Collisions in VR . . . . . . . . . . . . . . . . . . . . . . . . . . . 626

Marek Kopel and Bartłomiej Stanasiuk
Generation of Musical Scores from Chord Sequences Using

Neurodynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
Koji Masago, Mizuho Amo, Shun Nishide, Xin Kang, and Fuji Ren
Improving Variational Mode Decomposition-Based Signal Enhancement

with the Use of Total Variation Denoising . . . . . . . . . . . . . . . . . . . . . . . . . 649
Krzysztof Brzostowski and Jerzy Świątek
Machine Learning
Colored Petri Net Modeling for Prediction Processes in Machine Learning . . . 663
Ibuki Kawamitsu and Morikazu Nakamura
Enriching the Semantics of Temporal Relations for Temporal

Pattern Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Ryosuke Matsuo, Tomoyoshi Yamazaki, Muneo Kushima,
and Kenji Araki
Integer Weighted Regression Tsetlin Machines . . . . . . . . . . . . . . . . . . . . . . 686

Kuruge Darshana Abeyrathna, Ole-Christoffer Granmo,
and Morten Goodwin
Increasing the Inference and Learning Speed of Tsetlin Machines

with Clause Indexing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
Saeed Rahimi Gorji, Ole-Christoffer Granmo, Sondre Glimsdal,
Jonathan Edwards, and Morten Goodwin
Constrained Evolutionary Piecemeal Training to Design Convolutional

Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
Dolly Sapra and Andy D. Pimentel
Hierarchical Learning of Primitives Using Neurodynamic Model. . . . . . . . . . 722

Fusei Nomoto, Tadayoshi Yasuda, Shun Nishide, Xin Kang,
and Fuji Ren
Compressing and Interpreting SOM-Based Convolutional

Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 732
Ryotaro Kamimura
Data Management and Data Clustering
A Quality Assessment Tool for Koblenz Datasets Using

Metrics-Driven Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 747
Szymon Pucher and Dariusz Król
Contents xvii
Applying Cluster-Based Zero-Shot Classifier to Data Imbalance Problems . . . 759

Toshitaka Hayashi, Kotaro Ambai, and Hamido Fujita
Determining Sufficient Volume of Data for Analysis

with Statistical Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
Tanvi Barot, Gautam Srivastava, and Vijay Mago
A Fuzzy Crow Search Algorithm for Solving Data Clustering Problem . . . . . 782
Ze-Xue Wu, Ko-Wei Huang, and Chu-Sing Yang
Distributed Density Peak Clustering of Trajectory Data on Spark . . . . . . . . . 792

Yunhong Zheng, Xinzheng Niu, Philippe Fournier-Viger, Fan Li,
and Lin Gao
Pattern Mining
Parallel Mining of Partial Periodic Itemsets in Big Data . . . . . . . . . . . . . . . . 807

C. Saideep, R. Uday Kiran, Koji Zettsu, Cheng-Wei Wu,
P. Krishna Reddy, Masashi Toyoda, and Masaru Kitsuregawa
A Fast Algorithm for Mining Closed Inter-transaction Patterns . . . . . . . . . . . 820

Thanh-Ngo Nguyen, Loan T.T. Nguyen, Bay Vo,
and Ngoc Thanh Nguyen
TKE: Mining Top-K Frequent Episodes . . . . . . . . . . . . . . . . . . . . . . . . . . . 832

Philippe Fournier-Viger, Yanjun Yang, Peng Yang, Jerry Chun-Wei Lin,
and Unil Yun
TKU-CE: Cross-Entropy Method for Mining Top-K High Utility Itemsets . . . 846
Wei Song, Lu Liu, and Chaomin Huang
Mining Cross-Level High Utility Itemsets . . . . . . . . . . . . . . . . . . . . . . . . . 858

Philippe Fournier-Viger, Ying Wang, Jerry Chun-Wei Lin,
Jose Maria Luna, and Sebastian Ventura
Efficient Mining of Pareto-Front High Expected Utility Patterns . . . . . . . . . . 872

Usman Ahmed, Jerry Chun-Wei Lin, Jimmy Ming-Tai Wu,
Youcef Djenouri, Gautam Srivastava, and Suresh Kumar Mukhiya
Maintenance of Prelarge High Average-Utility Patterns

in Incremental Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 884
Jimmy Ming-Tai Wu, Qian Teng, Jerry Chun-Wei Lin,
Philippe Fournier-Viger, and Chien-Fu Cheng
xviii Contents
System Control, Classification, and Fault Diagnosis
Development and Research of a Terminal Controller for Marine Robots . . . . . 899

V. Pshikhopov and Boris Gurenko
A Machine Learning Approach for Classifying Movement Styles Based

on UHF-RFID Detections. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 907
Christoph Uran, Markus Prossegger, Sebastian Vock,
and Helmut Wöllik
Process Decomposition and Test Selection for Distributed Fault Diagnosis . . . 914
Elodie Chanthery, Anna Sztyber, Louise Travé-Massuyès,
and Carlos Gustavo Pérez-Zuñiga
Managing Situations with High Number of Elements in Group

Decision Making. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 926
J. A. Morente-Molinera, S. Alonso, S. Ríos-Aguilar, R. González,
and E. Herrera-Viedma
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 933

Natural Language Processing
Question Generation Through Transfer
Learning
Yin-Hsiang Liao and Jia-Ling Koh(B)
Department of Computer Science and Information Engineering,

National Taiwan Normal University,
Taipei, Taiwan, R.O.C.
[email protected], [email protected]
Abstract. An automatic question generation (QG) system aims to pro-

duce questions from a text, such as a sentence or a paragraph. Traditional
approaches are mainly based on heuristic, hand-crafted rules to trans-
duce a declarative sentence to a related interrogative sentence. However,
creating such a set of rules requires deep linguistic knowledge and most
of these rules are language-specific. Although a data-driven approach
reduces the participation of linguistic experts, to get sufficient labeled
data for QG model training is still a difficult task. In this paper, we
applied a neural sequence-to-sequence pointer-generator network with
various transfer learning strategies to capture the underlying informa-
tion of making a question, on a target domain with rare training pairs.
Our experiment demonstrates the viability of domain adaptation in QG
task. We also show the possibility that transfer learning is helpful in
a semi-supervised approach when the amount of training pairs in the
target QG dataset is not large enough.
Keywords: Question generation · Sequence-to-sequence model ·

Transfer learning
1 Introduction
Nowadays, teachers in school are far from the only way in which students can
get knowledge. Besides from the traditional education such as classrooms, there
are plenty of sources to be chosen, like massive open online courses (MOOCs)
or open educational materials. Sufficient and frequent quizzes help students get
better learning outcomes than just studying textbooks or notes [3,8]. However,
creating reasonable and meaningful questions is a costly task in both time and
money. The amount of related quizzes is not comparable with the amount of
growing online educational materials. Accordingly, it is worthwhile to build a
This research is partially supported by the “Aim for the Top University Project” of
National Taiwan Normal University (NTNU), sponsored by the Ministry of Education
and Ministry of Science and Technology, Taiwan, R.O.C. under Grant no. MOST 108-
2221-E-003-010.
c Springer Nature Switzerland AG 2020
H. Fujita et al. (Eds.): IEA/AIE 2020, LNAI 12144, pp. 3–17, 2020.
https://doi.org/10.1007/978-3-030-55789-8_1
4 Y.-H. Liao and J.-L. Koh
reliable automatic question generation (QG) system for educational purpose.

Meanwhile, the previous works are generally based on English [6,7,17–20]. In this
paper, we aim to create a system to utilize educational resources in traditional
Chinese for middle-school students.
In computational linguistics, there are two mainstream approaches for QG –
rule-based approaches and data-driven approaches. Rule-based approaches basi-
cally utilize grammars to generate texts. Ideally, the rule-based approaches can
generate good questions if researchers design rules well. However, creating such
a set of rules requires deep linguistic knowledge and most of these rules are
language-specific. In order to reduce the participation of experts, a data-driven
approach is required to alleviate the need of strong linguistic background for a
certain language. In addition, the proposed methods should fit our application
scenario in traditional Chinese.
A data-driven question generation system for educational purpose has dif-
ficulties as following. 1) Proper labeled data is insufficient. Training question
generation models need input texts, paragraphs or sentences, and correspond-
ing questions in supervised approaches. Quiz questions are relatively few with
respect to the educational content. Even if quizzes are abundant because of an
existing question bank, the annotated sentence-question pairs remain rare. 2)
The absence of an effective evaluation metrics. In previous works, the most pop-
ular evaluation metrics include BLEU [11] and ROUGE [9]. So far, there has
been no explicit and correct way to objectively evaluate whether the generated
questions of a model are good for the given content.
To materialize the aforementioned system, we mainly study how to apply a
supervised approach of transfer learning on a Seq2seq pointer network (PTN)
model for solving the QG task with rare sentence-question training pairs. We
also consider a semi-supervised approach combined with domain adoption. Our
main contributions are as follows:
1) We demonstrate the viability of domain adaptation in QG task.

2) We perform several fine-tuning strategies to construct a Seq2seq PTN model
for QG with a better performance.
3) We investigate the feasibility of semi-supervised fine-tuning for the Seq2seq
PTN model.
4) In addition to BLEU and ROUGE, we use an additional simple evaluation,
the proportion of generated questions with interrogatives, to evaluate the
generate questions.
5) The proposed framework demonstrates the possibility of applying the QG
system to school teaching materials.
This paper is organized as follows. In Sect. 2, we introduce the related works.
In Sect. 3, we describe the proposed approaches in detail. Section 4 shows per-
formance evaluation of the proposed methods and discusses the results of exper-
iments. In Sect. 5, we conclude this research and point out the future works.
Question Generation Through Transfer Learning 5
2 Related Works
Researchers have dealt with question generation by rule-based approaches in the
past [12]. These solutions depended on well-designed rules, based on profound
linguistic knowledge, to transform declarative sentences into their syntactic rep-
resentations and then generate interrogative sentences. [7] took an “overgenerate-
and-rank” strategy, which used a set of rules to generate more-than-enough
questions and leveraged a supervised learning method to rank the produced
questions. The rule-based approaches performed well on well-structured input
text. However, because of the limitation of hand-crafted rules, the systems failed
to deal with subtle or complicated text. In addition, these heuristic rule-based
approaches focused on the syntactic information of input words, most of which
ignored the semantic information.
Du et al. [6] first proposed a Seq2seq framework with attention mechanism
to model the question generation task for reading comprehension. Their work
considered context information both from a sentence and a paragraph. [18] pre-
sented another QG model with two kinds of decoders. By considering the types
of words including interrogatives, topic words, and ordinary words, the model
aimed to generate questions for an open-domain conversational system. To lever-
age more data with potentially useful information, the answer of a question was
considered in [15] and [17]. The work of Tang et al. showed that the QA and QG
tasks enhanced each other in their training framework.
The aforementioned QG works were remarkable; nevertheless, their successes
were inseparable with SQuAD, a relatively large, publicly available, general pur-
pose dataset. If there is a shortage of training data in a domain where we are
interested, e.g., teaching materials in middle school, it is difficult to train the
models to an acceptable level. For educational purpose, Z. Wang et al. [19] pro-
posed a Seq2seq-based model, QG-net, that captures the way how humans ask
questions from a general purpose dataset, SQuAD, and directly applied the built
model to the learning material, the OpenStax textbooks. Similarly, our study
aims to build a system for middle school education but there does not exist a
large-scale dataset with manually labeled sentence-question pairs. However, our
work mainly differs from QG-net in the following aspects. First, we focus on the
effectiveness of domain adaptation: we tune the proposed model by hundreds of
labeled pairs in our target domain, middle school textbooks. Second, since we
have labeled data in target domain, we therefore do quantitative evaluations,
which were unseen for the generated questions in the target domain of QG-net.
Furthermore, our study proposes a semi-supervised approach to leverage more
generated questions as training pairs to fine-tune the baseline model.
Recently, in natural language processing (NLP) community, there were var-
ious applications of the Seq2Seq model. A Seq2seq model typically consists an
encoder and a decoder. Most of the frameworks are implemented by RNNs. The
encoder looks through the input text as a context reader, and converts it to a
context vector with textual information. The vector is then decrypted by the
decoder as a question generator. The decoding procedure often take the atten-
tion mechanism [1,10] to generate a meaningful question corresponding to the
input text. [16] proposed the pointer network, a modification of Seq2seq, to deal
with the words absent in the training set. Their work was later used by Z. Wang
et al. [19] to point out which part of an input content is more possible to appear
in the output question. We will describe in detail how we apply the Seq2seq
model and its variations in Sect. 3.
Deep neural networks (DNNs) often benefit from transfer learning. In NLP,
transfer learning has also been successfully applied in tasks like QA [5], among
other things. [5] demonstrated that a simple transfer learning technique can be
very useful for the task of multi-choice question answering. Besides, the paper
showed that by an iterative self-labeling technique, unsupervised transfer learn-
ing is still useful. Inspired by [5], we performed experiments to investigate the
transferability of encoder and decoder learned from a source QG dataset to a
target dataset using a sequence-to-sequence pointer-generator network. The size
of the target dataset considered in our study is even smaller than that used in [5].
Although unsupervised transfer learning for QG is still a challenge, we showed
that transfer learning is helpful in a semi-supervised approach.
3 Methods
The proposed method consists of three part of processing: data preparation,

training baseline model, and domain adaptation.
3.1 Task Definition

S
Given an input sentence S = {wk }L
k=1 , where wk is a word in the input sentence
S with length LS . The goal of question generation is to generate a natural
question Q, to maximize P (Q|S, θ) [6].
Q
P (Q|S, θ) = t=1 P (qt |S, {qτ }τ =1 }, θ),
L t−1
(1)
where LQ denotes the length of output question Q, and qt denotes each word
within Q, respectively. Besides, θ denotes the set of parameters of a prediction
model to get P (Q|S, θ), the conditional probability of the predicted question
sequence Q, given the input S. A basic assumption is that the answer of the
generated question should be a consecutive segment in S. Accordingly, we mainly
consider how to generate the factual questions.
3.2 Data Preparation
For preparing the source domain data and target domain data, the data pre-
processing steps are required as follows.
Source Domain Data. In the DRCD dataset, each given data consists of
a triple (P , Q, A), where P is a paragraph, Q is a question from the given
paragraph, and A is the answer in the paragraph. The following processing is
performed to get a pair of input sentence S and the corresponding question Q.
1. Extract the sentence S that contains the answer A from the paragraph P .
2. Generate the sentence-question pairs (S, Q).
3. Segment the texts in each sentence-question pair via Jieba [14] .
The constructed dataset is denoted by DBDRCD .
Target Domain Data. We collected a dataset of sentences in textbooks of

junior high school in science subject and a dataset of multi-selection questions
from a question bank. Then the proper questions are selected and matched with
the sentences in the textbooks semi-automatically as the following:
1. Exclude the questions with a figure(s).
2. Exclude the questions with the form “which of the following is correct/wrong”
and keep the factual questions.
3. Match each remaining question Q and its answer A with each sentence S in
the textbook as follows:
– If A appears in the sentence S, generate a candidate sentence-question
pair (S, Q) and compute BLEU-4 scores [11] between S and Q.
4. Select the sentence-question pairs with BLEU-4 scores higher than a threshold
value and manually make the proper pairs remaining.
5. Exclude duplicating sentence-question pairs.
6. Perform word segmentation on the texts of the pairs via Jieba.
3.3 Baseline Model

The baseline neural model used in this study is the pointer network model. We
describe details of the model as follows.
Our encoder and decoder are implemented by a bidirectional Gated Recurrent
Unit (bi-GRU) [1] and a GRU, respectively [4]. The words in S, i.e. wi , are
inputted to an embedding layer, and then inputted into the encoder one-by-one.
Then a sequence of hidden states hi of the encoder is produced.
In the Seq2seq model, the attention distribution indicates the weights of
hidden states in GRU encoder cells for predicting the result of decoder. On
each step, the decoder receives the word embedding of the previously generated
word qt−1 and produces the hidden state of decoder at step t, hdt . The attention
distribution aet is calculated as:
eti = hTi Wattn hdt + battn , (2)
aet = sof tmax(et ). (3)
states of encoder, hi , are weighted summed to get the context

Then the hidden
vector: cet = i aeti hi .
Moreover, we apply the intra decoder attention mechanism [20], where the
attention distribution adt for each decoding step t is calculated as:
T
edtt = hdt W dattn hdt + bdattn , (4)
adt = sof tmax(edt ), (5)

where hdt andhdt denote the hidden states of decoder at step t and the previous
step t , respectively. Accordingly, another context
d vector cdt is generated by the
intra decoder mechanism as follows: ct = d
t att ht . After concatenating ht
d d
with ct and ct , denoted by , an output layer generates the token by computing

e d
the following probability distribution:
Pvoc (qt ) = sof tmax(Wout [hdt cet cdt ] + bout ). (6)
Fig. 1. The framework of pointer-generator network.
Pointer Network (PTN) is a variation of the Seq2seq model. It calculates

the output word probabilities as a weighted sum of two probabilities, one comes
from the output of Seq2seq model and another comes from the attention weights
of input sentence. As shown in Fig. 1, the proportion of the two probabilities are
controlled by the tunable parameter Pgen .
Pgen = σ(Wu [hdt cet cdt ] + bu ). (7)

The PTN will create an extra dictionary for the unseen words occurring in the
input sentence. Those words will be probably chosen to be the output when their
probabilities from the attentions weights of input are large. PP T N (qt = xi ) = aeti ,
where xi is in S. Finally, P (qt ) = Pgen ×Pvoc (qt )+(1−Pgen )×PP T N (qt ). By the
usage of point network, this model can effectively deal with out-of-vocabulary
problem. In order to prevent repetition in the generation model, coverage mech-
anism [13] is applied. A coverage value covt is computed, which is the sum of
attention distributions over all previous decoder time steps and contributes to
the loss:
covt = aeti , (8)
i<t

covloss = min(aeti , covt ). (9)
i<t
The total loss is losst = logP (w∗t ) + λ × covlosst , where P (w∗t ) is the negative
log likelihood (NLL) of the target word qt for the time step t. λ is a given hyper-
parameter to control the weight
of coverage loss. The overall loss for the whole
sequence is: loss = 1/LS × t losst . Note that we did not put Chinese characters
into GRU directly. Instead, we first built a dictionary of frequent words from the
word segmentation results of the training set. For each word in the dictionary,
if it is in the vocabulary of fastText [2], its corresponding pre-trained word
embedding is loaded.
In the training process of a Seq2seq model, the inference of a new token
is based on the current hidden state and the previous predicted token. A bad
inference will then make the next inference worse. This phenomenon is a kind
of error propagation. D. Bahdanau et al. [1] thus proposed a learning strategy,
named teacher forcing, to ease the problem. Instead of always using the gener-
ated tokens, the strategy gently changed the training process from fully using
the true tokens, toward mostly using the generated tokens. This method can
yield performance improvement for sequence prediction tasks such as QG. In
the proposed model, we guide the training by 0.75 at beginning, and decay the
ratio by multiplying 0.9999 after each epoch.
3.4 Domain Adaptation
Supervised Domain Adaptation. We take the supervised transfer learning

approach as the following three steps:
1. Given an epoch number epon to train the PTN on the source domain, which
contains abundant training data, and save the model of each epoch until reach
the number epon.
2. Select a best model according to a selection strategy as the base model Mb .
In the experiment, we choose the model with the highest average BLEU-4 on
the validation set of target domain.
3. Fine-tune the model Mb . That is, we initialize another training process with
learnt parameters of the model Mb on the dataset set of target domain.
The PTN consists of the following layers: the embedding layer for
encoder/decoder, the bi-GRU layer and attention network for encoder, the GRU
layer and attention network for decoder, the output layer for question genera-
tion, and the parameter for computing Pgen . We try various strategies of domain
adoptions by freezing or retraining some layers of the model. The details of var-
ious fine-tuning strategies are described in the experiments.
Semi-supervised Domain Adaptation. Algorithm 1 shows our basic semi-

supervised approach.
Algorithm 1: Semi-supervised QG Domain Adaptation

Input: Source dataset DBsource : (sentence, question) pairs;
Target datasets DBtarget : DB ltarget :(sentence, question) pairs;
DB ul
target : sentences; Number of training iterations n.
Output: QG model M*
1. Pre-train QG models on the source dataset.
2. Select a best Pre-trained model Mb .
3. Repeat
4. For each sentence S in the target dataset DB ul target , use Mb to generate
its question Q.
5. Combine the predicted question Q with S to generate a training pair of
DB ul
target .
6. Fine-tune Mb according to DB ltarget ∪ the training pairs of DB ul target .
7. Until reach the nth iteration
We also proposed an enhanced version of semi-supervised QG domain adapta-
tion. Instead of applying Mb to predict the questions of DB ul
target directly, Mb is
fine-tuned by DB target after step 2. The model with the highest average BLEU-4
l
on the validation set of target domain is then selected to perform the iteration
from step 3 to 7.
4 Performance Evaluation
4.1 Experiments Setup
For training the base model, we used the DBDRCD , which is an open domain tra-
ditional Chinese machine reading comprehension (MRC) dataset. The dataset
contains 10,014 paragraphs from 2,108 Wikipedia articles and 30,000+ ques-
tions generated by annotators. We excluded those sentence-question pairs over
80 and 50 words, respectively, so 26,175 pairs are extracted. Moreover, the dictio-
nary of vocabulary contains the ones with frequency no less than 3. Therefore,
the vocabulary size of DBDRCD is 28,981. In the target domain, the dataset
DBtextbook contains 480 labeled pairs. We applied random sampling to separate
the data into train/test/validation sets in the proportion 7/2/1. Accordingly,
there are 336 pairs with vocabulary size 787 in the training set, 48 pairs for
validation, and 96 pairs for testing. The effectiveness of transfer learning is eval-
uated by the model’s performance on the test sets of target domain. For the
training set consisting of both DBDRCD and DBtextbook , the pairs/vocabulary

size is 26,511/29,201. It implies that there are 220 words in the vocabulary of
DBtextbook but not in the vocabulary of DBDRCD . For all the experiments,
our models have 300-dimensional hidden states and 300-dimensional pre-trained
word embedding. We set learning rate to be 0.0001, teacher forcing rate 0.75,
batch size 32, and use Adam as the optimizer.
We report the evaluation results with the following metrics.
– BLEU-4 [11]. BLEU measures precision by how much the words in prediction
sentences appear in reference sentences. BLEU-4 measures the average for
calculation on 1-gram to 4-gram.
– ROUGE- L [9]. ROUGE-L measures recall by how much the words in reference
sentences appear in prediction sentences using Longest Common Subsequence
(LCS) based statistics.
– Ratio of interrogatives, denoted by ‘?%’. In Chinese, there are various inter-
rogatives corresponding to the 5W1H, i.e. ‘Who, Why, When, Where, What,
and How’, questions. A question with an interrogative usually shows the
explicit target of a question instead of a yes/no question, which is a ques-
tion with higher quality. Accordingly, we measure how many percentage of
sentences in the generated questions having one of the interrogatives.
4.2 Experiment Results

Experiment 1. In this experiment, three QG models, denoted as M1 , M2 , and
M3 , are constructed by varying the training dataset. Moreover, for each model,
two versions are trained by fixing or not fixing the pre-trained embedding layer
as shown in Table 1. The I/O Voc. denote the input and output vocabulary
of the model. The evaluation results of Experiment 1 are shown in Table 2.
The results show that, for the models constructed by training DBDRCD or
DBT extbook only, fixing the pre-trained embedding layer gets the better result
on the test data of DBT extbook than fine-tuning the embedding layer. Besides,
the model trained by DBDRCD (Model M1 ) outperforms the model trained
directly on the target dataset DBtextbook (Model M2 ). It indicates that the lim-
ited size of the target dataset could not provide enough data for the complex
structure of the PTN generation model. Moreover, by using the training data of
DBDRCD ∪ DBT extbook could not solve the problem properly, because the ratio
of DBT extbook to DBDRCD is tiny. It implies the necessarily of transfer learning.
Table 1. The descriptions of the constructing models of Experiment 1.
Model I/O Voc. Training dataset Fixed emb

Model M1 /M1 DBDRCD DBDRCD Yes/No
Model M2 /M2 DBT extbook DBT extbook Yes/No
Model M3 /M3 DBDRCD ∪ DBT extbook DBDRCD ∪ DBT extbook Yes/No
Table 2. The evaluation results of Experiment 1.
Fix embedding layer Fine-tune embedding layer

Model BLEU-4 ROUGE L ?% Model BLEU4 ROUGE L ?%
Model M1 0.344 0.430 0.0625 Model M1 0.295 0.401 0.115
Model M2 0.259 0.388 0.208 Model M2 0.193 0.336 0.083
Model M3 0.299 0.414 0.052 Model M3 0.314 0.418 0.0625
Experiment 2. The purpose of Experiment 2 is to demonstrate the effectiveness

of domain adaptation. In the following experiments, the best model in M1 ’s
training process is selected as Mb . Various versions of domain adaptation are
designed by deciding whether transferring the learnt parameters of layers in Mb
related to the input/output vocabulary.
Table 3. The descriptions of the constructing models of Experiment 2.
Model I/O Voc Retraining layers

Model M4 DBDRCD Embedding for encoder & decoder, Out
Model M5 DBDRCD Out
Model M6 DBDRCD None
Model M7 DBDRCD /DBT extbook Embedding for decoder, Out
Model M7∗ DBDRCD /DBT extbook Out
Model M7a DBDRCD /DBT extbook + ‘?s’ Embedding for decoder, Out
Model M8 DBT extbook Embedding for encoder & decoder, Out
Model M8a DBT extbook + ‘?s’ Embedding for encoder & decoder
Table 4. The evaluation results of Experiment 2.
Model BLEU-4 ROUGE L ?% Model BLEU4 ROUGE L ?%

Model M1 0.344 0.430 0.0625 Model M7 0.340 0.456 0.365
Model M4 0.352 0.438 0.010 Model M ∗7 0.305 0.426 0.0625
Model M5 0.366 0.443 0.010 Model M7a 0.355 0.470 0.177
Model M6 0.426 0.526 0.323 Model M8 0.438 0.535 0.302
Model M8a 0.447 0.545 0.354
Five strategies for constructing the models by fine-tuning M1 , to get mod-

els M4 to M8 , are described as Table 3. The retraining layers means the
learnt parameters of the layers in Mb were not loaded, which are retrained
Another random document with
no related content on Scribd:
8
In speaking of Doris, Herodotus says: Ή δὲ χώρη αὕτη ἐστὶ
μητρόπολις Δωριέων τῶν ἐν Πελοποννήσῳ. That the land had a
Dorian population in the fifth century b.c. is undoubtedly the case;
but its claim to be metropolis of the Dorians of the south was in all
probability set up by the Spartan authorities, as affording a
convenient pretext for interference in Greek affairs north of Isthmus.
It is probable that this corner of Greece, of which the Malian plain
was the centre, contained patches of various peoples which had in
different ages traversed the peninsula, or which had been driven into
its mountain fastnesses by the passage of invaders:⁠—Dorians,
Œtæans, Trachinians, etc., were probably such remains of larger
tribes.
9
Herodotus seems from his language to assume (viii. 31) that the
whole army had come south by the Dorian route. That is, either a
mistake; or, more probably, the impression his language gives is due
to a mere omission. Few details are given of any part of the route of
the army.
0
The position at Delphi, from a military point of view, is by no means
weak, provided Amphissa be occupied, and the great pass from the
north be thus closed. Under those circumstances, unless the
assailant is in a position to land troops at the head of the Krissæan
gulf, the only line of attack is along this easily defensible path from
the west. It is imaginable that Xerxes, knowing it to be an open town,
under-estimated the difficulty of its capture.
Pogon is an almost land-locked harbour between the island of
Kalauria and the mainland.
2
A Comparison of the Lists of Vessels at Artemisium and

Salamis respectively.
T. = trireme; P. = pentekonters.
Artemisium. Salamis.
{127 T.
Athenians (some Platæans in crews at
53 T. later} 180 T.
Artemisium)
180 T.
Corinthians 40 T. 40 T.
Megareans 20 T. 20 T.
Chalkidians in Athenian ships 20 T. 20 T.
Æginetans 18 T. 30 T.
Sikyonians 12 T. 15 T.
Lacedæmonians 10 T. 16 T.
Epidaurians 8 T. 10 T.
Eretrians 7 T. 7 T.
Trœzenians 5 T. 5 T.
Styreans 2 T. 2 T.
Keians 2 T.; 2 P. 2 T.; 2 P.
Opuntian Locrians 7 P. —
Hermionians — 3 T.
Ambrakiots — 7 T.
Leukadians — 3 T.
Naxians — 4 T.
Kythnians — 1 T.; 1 P.
Krotonians — 1 T.
Malians — 2 P.
Siphnians — 1 P.
Seriphians — 1 P.
366 T.;
324 T.; 9 P.
7 P.
Æschylus gives 310 as the number of the Greek fleet. Valuable as
is the testimony of the poet with regard to those incidents in the
battle which he observed as an eye-witness, his evidence on the dry
question of numbers is not likely to be exact.
3
She did, indeed, send sixty vessels, to observe, so said the patriot
Greeks, how the war went, but not with any intention of taking part
therein. The Corcyræans’ own excuse for their non-participation was
that their fleet had been unable to round Malea.
4
The mistake may be that of a manuscript copyist; but such mistakes
are so common in the text of Herodotus, that they afford strong
ground for supposing that the historian was, like the men of his time,
inaccurate in numerical calculations. The mistake may be in the
Paus. ii. 29. 5. detailed list. Pausanias implies that the Æginetan
contingent was superior in numbers to that of the
Corinthian, that is to say, more than forty. If the number were forty-
two, the total given by Herodotus would be correct; and it is
noticeable in this reference that he himself, in speaking of the
H. viii. 46. number of ships which Ægina supplied, says: “Of the
islanders the Æginetans supplied thirty; they had indeed
other ships manned; but with these they were guarding their own
country; but with the thirty best sailers they fought at Salamis.”
5
By Professor J. W. Bury.
6
Macan, Herod, iv., v., vi., “Athens and Ægina.”
7
Note on the Reference to Siris in Themistocles’ Speech.—The
reference to Siris inevitably suggests that this reported passage in
Themistocles’ speech is an invention of later date arising from the
colonization of Thurii in or about 443. The rapid growth of Athenian
trade in the earlier part of the fifth century, and its peculiar
development along the western route, render it possible, however,
that an idea of settlement on or near the deserted city of Sybaris
may have been long anterior to the actual settlement, and may have
been mooted even before 480. If Plutarch is to be believed,
Themistocles had direct relations with Corcyra, and gave the name
of Sybaris to one, and the name of Italia to the other of his daughters
(Plut. Them. 32).
8
H. viii. 74. τέλος δὲ ἐξερράγη ἐς τὸ μέσον. Cf. also Diod. xi. 16, ad
fin.
9
This Council of War must have been held on the morning of the day
preceding the battle. It lasted, in all probability, several hours, and, if
so, this would indicate the afternoon as the time at which Xerxes
received the message of Themistocles. On this point, then, the
indications in the narrative of Æschylus and Herodotus are in
agreement.
0
Plut. Them. also mentions the same name; but the testimony is
probably dependent on that of Herodotus.
There is a curious triangular concord at this point in the history.
Diodorus says that the Egyptian contingent was sent to block the
strait towards the Megarid (xi. 17).
Plutarch says 200 vessels were sent to close the passage round
Salamis (Them. 12).
Herodotus mentions that the Egyptian contingent numbered 200
(vii. 89).
2
This would account for the fact implied by Æsch. Pers. 400: the two
fleets when they started their movement were not in sight of one
another, though, very shortly after the movement began, the Persian
fleet was visible to the Greeks. The latter would first catch sight of it
after it rounded the Kynosura promontory and the island.
3
Cf. Arist. 8, where the revocation is said to have taken place τρίτῳ
ἔτει after the sentence.
4
Cf. Stein’s brief note on the translation of the words στὰς ἐπὶ τὸ
συνέδριον in H. viii. 79.
5
In so far as I know, this last very important point was first raised by
Prof. J. B. Bury in an article in the Classical Review on “Aristides at
Salamis.”
6
This is Professor Bury’s suggestion. It is open to the objection that
Herodotus expressly mentions the arrival of this vessel (H. viii. 83)
immediately before the battle began. But this objection is not by any
means insuperable. It is much more probable, under the
circumstances, that Herodotus made a mistake as to the time of its
arrival, than that it managed at the time he mentions to force its way
through the blockading fleets at either end of the strait.
7
It would seem as if it were a description of this movement, taken
from his notes on, or sources of information for, the details of the
battle, which Herodotus has used by mistake in describing the
movement of the Persian fleet during the night. He has, of course,
intensely confused the original description by reading into it what he
knew to be the object of that night-movement—the surrounding of
the Greek fleet by blocking the issues both to east and west of it; but,
eliminating this motive from his description, it is possible to see that
in its original form it must have resembled very closely the
description of the advance of the Persian fleet which has been drawn
from the details which Æschylus and Diodorus give.
H. viii. 76. “The west wing put out and made a circling movement
towards Salamis.” It has been already pointed out that by “west
wing” Herodotus evidently means, not the west wing in the original
formation, but the west wing when the fleet had completed the
movement, and had taken up the position which he imagined it to
have assumed when the movement was complete. This “west wing”
would be the east wing in the original position. That it cannot have
been the original west wing has been pointed out in a previous note.
If this correction be made, Herodotus’ language in describing this
movement is peculiarly applicable to the movement of that part of the
Persian fleet which entered the strait by the channel east of
Psyttaleia—ἀνῆγον κυκλούμενοι πρὸς τὴν Σαλαμῖνα; and the
applicability becomes still more striking in view of the evidence,
which will be given later, that this wing of the Persian fleet got in
advance of the other.
The left wing, which would use the channel west of Psyttaleia, is
equally referred to in the words: “Those about Keos and Kynosura
put out in order,” to which he adds, in accordance with his knowledge
that part of the object of the night-movement was the blocking of the
straits, “And they occupied the whole strait as far as Munychia with
their ships.”
8
This phenomenon of the morning wind is very common in the Greek
seas. It will be remembered that Phormio based his tactics in his first
battle with the Corinthian fleet just outside the Corinthian gulf on its
occurrence. I have experienced it there; and on the three occasions
on which I have been through the Strait of Salamis, once in the
summer of 1895, and twice in the summer of 1899, I have
experienced it on each occasion. It began in all three cases quite
suddenly, a little before seven in the morning, blowing from the west,
right down that part of the strait south of Ægaleos. It was extremely
violent while it lasted, though it did not raise a dangerous sea. To the
inexperienced it gave the impression that it meant the beginning of a
very windy day. On two occasions it ceased about 8.30, on the other,
shortly after nine, and the dead calm by which it had been preceded
ensued once more.
9
As is shown by the presence of an Attic vessel opposite the Persian
left, where her ships must almost certainly have been.
0
Cf. Æsch. Pers. 724,—Ναυτικὸς στρατὸς κακωθεὶς πεζὸν ὤλεσε
στρατόν. Thuc. i. 73, 5.—Νικωθεις γὰρ ταῖς ναυσίν ὡς οὐκέτι αὐτῷ
ὁμοίας οὔσης της δυνάμεως κατὰ τάχος τῷ πλέονι τοῦ στρατοῦ
ἀπεχώρησεν.
Modern historians have taken this account of the intended or
attempted construction of the mole too seriously. It has been pointed
out, for instance, that the only point in the strait east of the bay of
Eleusis at which it could possibly be carried out, is at the narrows
where the island of St. George contracts the width of the channel,
and that it is impossible that, under the circumstances as they stood,
Xerxes should have been able to bring vessels to that part of the
strait. But Herodotus never attempts to give the impression that the
operation was ever undertaken seriously; he makes it plain, indeed,
that it was not. If that were so, and it was merely designed to give
the Greeks a wrong impression, it did not in the least matter whether
it was made at a possible or impossible point. Ktesias, Pers. 26, and
Strabo, 395, say that the mole was begun before the battle. This
would imply that a serious attempt was made to construct it. The
notorious unreliability of Ktesias, and the lateness of Strabo’s
evidence, render this account of the matter unworthy of
consideration.
2
H. viii. 103. Λέγουσα γὰρ ἐπετύγχανε τὰ πὲρ᾿ αὐτὸς ἐνόεε.
3
Οὐδεμία συμφορὴ μεγάλη ἔσται σεό τε περιεόντος καὶ ἐκείνων τῶν
πρηγμάτων περὶ οἶκον τὸν σόν.
4
It has been suggested that the real intention was to induce the
Ionians to revolt. The behaviour of this contingent in the recent battle
was not calculated to encourage such a plan, conceived within a few
days of the actual fight.
5
Ἐπείτε οὐκ ἐπαύετο λέγων ταῦτα ὁ Τιμόδημος, etc.
6
May it not be suggested that some archæologist acquainted with the
extant remains of Phœnician Carthage might confer a distinct
service on history by examining the structures at Agrigentum which
date from this period? The workman as well as the designer must
have set his mark there.
7
It has already been remarked that his description of Thermopylæ is
that of a traveller coming from the north—“from Achaia”—as he
himself says.
8
Herodotus himself (ix. 8) takes this view of the matter. He implies
that the Spartans did not care whether the Athenians medized or not
after the wall was completed. It is quite out of the question, however,
to suppose that the Spartans could have regarded with equanimity
the possible transference of the Athenian fleet to the Persian side.
They had the experience of Artemisium and Salamis to guide them.
9
It is sometimes assumed from H. vii. 229, that the usual quota was
one helot to each hoplite; but a more probable interpretation of that
passage is that the reference is to the personal armed servant who
accompanied each hoplite to war, and that it cannot be deduced
therefrom that the body of these formed the whole number of the
helots present on an ordinary occasion.
Modern criticism of the impossibility of despatching so large a
force unknown to the Athenian embassy is not convincing. We do
not know the place at which it gathered. It is extremely likely that a
large number of helots were drawn from Messenia, and joined the
army at Orestheion, where the great route from Messenia meets the
route from Sparta by way of the valley of the Eurotas.
0
His departure from the Isthmus is ascribed by Herodotus to the fact
that when he was sacrificing ἐπὶ τῷ Πέρσῃ an eclipse of the sun took
place. This eclipse has been calculated to have occurred on the 2nd
of October, 480. If so, it would be about the time of the Persian
retreat from Attica after Salamis, and Stein’s conjecture that the
sacrifice had something to do with a plan to harass the Persian
retreat, has a certain amount of probability in its favour.
If Sparta had been careless as to whether Athens medized or not,
she might, probably would, have despatched troops to the Isthmus at
an earlier date. But if she was waiting until pressure of
circumstances forced Athens to adopt Peloponnesian views as to the
line of defence, then the delay is accounted for. Had her army been
at the Isthmus when Mardonius advanced into Bœotia, the
Athenians would certainly have called upon it to carry out the
agreement, and march to the northward of Kithæron. In that case the
Spartan government would have been obliged either to comply, or,
by a refusal, to show in the most unmistakeable manner possible the
war policy which it intended to adopt.
2
I was, I confess, surprised to find in August, 1899, that, in spite of
the excellent road to Megara from Bœotia by the way of Eleusis, the
track on the old line of the Platæa-Megara road is still largely used.
3
A road has been constructed through it in recent years, running from
Kriekouki on the Bœotian side to Villa on the south of the range.
4
I am inclined to think that the site of Skolos is that which Leake, and
others following him, have identified with Erythræ. Paus. ix. 4, 3,
says that if before crossing the Asopos river on the road from Platæa
to Thebes, you turned off down the stream, and went about forty
stades, i.e. four and three-quarter miles, you came to the ruins of
Skolos. This would place it not far east of the road from Thebes to
Dryoskephalæ. He speaks of Skolos in another passage as a village
of Parasopia beneath Kithæron, a rugged place, and δυσοικητός.
That seems to preclude the idea of its being near the river, which
traverses alluvial lands at this part of its course. The ruins identified
by Leake as Erythræ cannot belong to that town if the testimony of
Herodotus and Pausanias is accurately worded. This point will be
discussed in a later note. In actual fact, however, the exact site of
Skolos is very difficult to determine. My main reason for suggesting
that it stood where Leake places Erythræ is that those ruins are the
only ruins in the neighbourhood indicated by Pausanias, and are
certainly not the ruins of Erythræ.
5
It is necessary to pursue so obvious a line of argument, because, for
some incomprehensible reason, modern historians have thought it
right to judge of the plans of these able Persian commanders as
though they were dictated by no higher considerations than such as
might occur to an untutored savage.
6
The weakness of this line in case of attack from the north was
conclusively shown twenty years later in the manœuvres which led
to the battle of Tanagra.
7
It is almost certain that an ancient road from Eleusis followed the
eminently natural line taken by the modern road from Eleusis to
Eleutheræ. There was also, in all probability, a route from Athens to
Eleutheræ which did not enter Eleusis at all, but, branching from the
Sacred Way near the Rheitoi after traversing the low pass through
Mount Ægaleos, went up the Thriasian plain and joined the road
from Eleusis among the low hills of Western Attica.
8
These ridges will be found numbered 1, 2, 3, 4, 5, 6, in the
accompanying map.
9
Marked A 6 in the map.
0
Called in the map, for purposes of distinction, the Asopos ridge, the
Long ridge, and the Plateau.
Those of the streams marked A 4 and A 5 on the map.
2
During my stay at Kriekouki, in December ’92–January ’93, the
rainfall was at times extraordinarily heavy. Nevertheless, I had not on
any occasion the slightest difficulty in crossing any of the streams,
and it was not even necessary to get wet in so doing. On one
occasion also I happened to be following the line of one of the
watercourses leading to the Œroë amid a downpour of rain such as
we rarely see in England, which had been going on with more or less
continuity for the previous fourteen hours; and yet, as I descended
the brook towards the plain the water became less and less until, on
the plain, there was no water running in the stream bed.
3
Pausanias knew the roads through these two passes.
(1) Platæa-Athens road.
He says (xi. 1, 6) that Neokles, the Bœotarch, in his surprise of
Platæa in the year 374, led the Thebans οὐ τὴν εὐθεῖαν ἀπὸ τῶν
Θηβῶν τὴν πεδιάδα, τὴν δὲ ἐπὶ Ὑσιὰς ἦγε πρὸς Ἐλευθερῶν τε καὶ
τῆς Ἀττικῆς.
There will be occasion to show that Hysiæ was in all probability a
small place, on a site just outside the southernmost end of the village
of Kriekouki. It was therefore at the eastern side of the opening of
the valley through which the road from Platæa to Athens passed.
The remains of that ancient road are, however, at the other side of
the valley opening; and, therefore, Hysiæ was not upon it. Probably,
however, down the valley came a track which is still used, and which,
after passing through the village of Kriekouki, goes due north to
Thebes in a line parallel to the main road from Dryoskephalæ. This
would be the road which Pausanias here mention. It would, in
entering the valley to the pass, go close to this site of Hysiæ. Of the
identity of this site it will be necessary to speak in a later note.
In 379, after the revolution in Thebes (X. H. v. 4, 14), the Spartans
despatched Kleombrotos with a force to Bœotia. As Chabrias, with
Athenian peltasts, was guarding “the road through Eleutheræ,” he
went, κατὰ τὴν εἰς Πλαταιὰς φέρουσαν.
This is almost certainly the Platæa-Athens pass. Kleombrotos
probably did not discover that the Dryoskephalæ pass was guarded
until he got to Eleutheræ. After doing so he turned to the left and
made his way through the Platæa-Athens pass, exterminating a
small body of troops which attempted to defend it.
(2) The Platæa-Megara road.
Pausanias (ix. 2, 3) says, Τοῖς δὲ ἐκ Μεγάρων ἰοῦσι πηγή τέ ἐστιν
ἐν δεξιᾷ καὶ προελθοῦσιν ὀλίγον πέτρα· καλοῦσι δὲ τὴν μὲν
Ἀκταίωνος κοίτην.
In the previous sentence he has expressly spoken of the road from
Eleutheræ to Platæa. The Megara road is therefore a different road.
The κοὶτη Ἀκταίωνος can, I think, be determined with sufficient
certainty at the present day. It is on the top of a low cliff, probably the
πέτρα mentioned, overhanging the sources of the stream O 3. Near
the foot of the cliff is an ancient well, known in Leake’s time as the
Vergutiani Spring.
4
Ἐπὶ τῆς ὑπωρέης τοῦ Κιθαιρῶνος.
5
The site of Erythræ.
Colonel Leake identified it with certain ruins which are found at the
foot of the mountain slope several miles east of the road from
Dryoskephalæ to Thebes. The available evidence seems to me to be
strongly against this view.
(1) The traditional site is where I have placed it, though I am afraid
that but little stress can be laid on traditions in modern Greece.
(2) Its comparatively frequent mention by Greek writers seems to
indicate that, though a small place, its position was of some
importance. If Leake’s view be correct this cannot have been the
case. If it were where I believe it to have been, it would be at the
northern exit of one of the most important passes in Greece. There is
an ancient φρουρίον on the bastion of Kithæron to the east of the
site. Its remains are so scanty, however, that they do not afford any
clue as to its date.
(3) There are remains of ancient buildings on the site. There are
also remains of an ancient well, besides which is a heap of stones,
from which two stones were obtained a few years ago with
inscriptions showing them to have belonged to a temple of
Eleusinian Demeter. Pausanias mentions so many temples in the
neighbourhood dedicated to that deity, that the discovery contributes
but little to the identification of the site. I was informed at Kriekouki
last year (August, 1899) that those particular stones were known to
have been originally discovered on another site. As neither my
informant nor any one else could tell me whence, why, or by whom
they were removed, I did not place much credence in the report.
(4) Pausanias says (ix. 2, 1), Γῆς δὲ τῆς Πλαταιίδος ἐν τῶ
Κιθαιρῶνι ὀλίγον τῆς εὐθείας ἐκτραπεῖσιν ἐς δεξιὰν Ὑσιῶν καὶ
Ἐρυθρῶν ἐρείπιά ἐστι; and further on (ix. 2, 2), he says, referring to
the road of which he is speaking: αὕτη μὲν (i.e. ὅδος) ἀπ’
Ἐλευθερῶν ἐς Πλάταιαν ἄγει. The road referred to is of course the
Athens-Platæa road, on which he is travelling towards Platæa. Can
any one suppose that Pausanias would have used the expression
quoted, especially the word ὀλίγον, had the ruins of Erythræ, as
Leake conjectured, lain some three and a half miles away from the
nearest point of this road, and hidden from it, moreover, by the great
projecting bastion of Kithæron, which is shown at the south-east
corner of the accompanying map?
Leake quotes Thucydides (iii. 24), who says that the two hundred
and twelve fugitives from Platæa first took the Thebes road in order
to put their pursuers off the scent, and then turning, ᾔεσαν τὴν πρὸς
τὸ ὄρος φέρουσαν ὁδόν ἐς Ἐρύθρας καὶ Ὑσιάς, καὶ λαβόμενοι τῶν
ὀρῶν διαφεύγουσιν ἐς τὰς Ἀθήνας. Meanwhile the pursuers were
searching the road along the ὐπωρέη. This last road would lead the
pursuers near the site where I conjecture Hysiæ to have stood, and
the objection may be raised that it is unlikely that the fugitives would
have gone to a place close to the road along which they could see
the pursuers were searching for them. It is, however, to be remarked
that Thucydides does not say that they went to either Erythræ or
Hysiæ. Had he intended to imply this he would have mentioned
those places in their proper order, Hysiæ first and Erythræ second.
Whenever he refers to the actual course taken by a body of men, or
by a fleet, he invariably mentions the places touched at or arrived at
in their geographical order. Vide Th. ii. 48, 1; ii. 56, 5; ii. 69, 1; iv. 5,
2; vii. 2, 2; vii. 31, 2.
The passage seems perfectly comprehensible and in accord with
the hypothesis which I put forward with respect to the positions of
Hysiæ and Erythræ. These fugitives, turning from the Platæa-
Thebes road, took the track which in modern times leads from
Pyrgos to Kriekouki, and which in ancient times would be the road
from Thespiæ to Hysiæ, Erythræ, and the passes. They did not go to
but towards those places, making in reality for those high rugged
bastions to the north-east of the pass of Dryoskephalæ.
But, after all, Pausanias’ words in the passage quoted dispose
effectively of Colonel Leake’s site. He would not have described a
place twenty-five stades away from the road as a short distance to
the right of it.
(5) Herodotus (ix. 15) speaks of the Persian camp as ἀρξάμενον
ἀπὸ Ἐρυθρέων παρὰ Ὑσιάς, κατέτεινε δὲ ἐς τὴν Πλαταίιδα γῆν.
These words merely show that Erythræ was east of Hysiæ.
(6) Perhaps one of the strongest pieces of evidence is Herodotus’
statement that the first Greek position was “at Erythræ.” Is it
conceivable that the Greek force, especially in its then state of
feeling with regard to the Persians, would be likely, after issuing from
the pass of Dryoskephalæ, to turn east along Kithæron, leave the
pass open, and take up a position with their backs to a part of the
range through which there was no passage of retreat?
(7) We are told later that their reason for moving to their second
position was the question of water-supply. This accords with the
present state of the locality about the traditional Erythræ. The
streams in that neighbourhood have but little water in them in the dry
season.
(8) The ground in this neighbourhood accords peculiarly with the
description given by Herodotus of the first engagement.
6
Marked ridges 1, 2, 3, 4, in the map.
7
These positions will be found marked upon the accompanying map.
It is necessary, however, to explain the evidence on which they are
determined.
8
The details of the contingents given by Herodotus are:⁠—
Lacedæmonians—
Spartans 5000
Periœki 5000
Helots 35,000
Tegeans 1500
Corinthians 5000
Potidæans 300
Orchomenians (Arcadia) 600
Sikyonians 3000
Epidaurians 800
Trœzenians 1000
Lepreans 200
Mykenæans and Tirynthians 400
Phliasians 1000
Hermionians 300
Eretrians and Styreans 600
Chalkidians 400
Ambrakiots 500
Leukadians and Anaktorians 800
Paleans from Kephallenia 200
Æginetans 500
Megareans 3000
Platæans 600
Athenians 8000
Miscellaneous light-armed troops 34,500
Total 108,200
9
I.e. A 1. In the days before scientific survey there was frequently the
utmost confusion with regard to the application of names to the head
streams of main rivers. This generally took the form of applying the
name of the main stream to several of its feeders. The tendency of
the local population was to apply the well-known name to that upper
tributary which was in their immediate neighbourhood, and was
therefore best known to them. Examples of this are frequent in
England; the upper waters of the Thames are a case in point. In
early sketch maps it will be found that the name Thames is applied
with the utmost diversity to the head streams of the river, and even a
tributary so far down as the Evenlode is sometimes given the name
of the main river. This is, I fancy, what has taken place with regard to
the Asopos. The Platæans, with whom Herodotus must have come
in contact in the course of his visit to the region, called this stream, A
1, by the name of the main river, and consequently “Asopos” in
Herodotus is to be understood to mean this stream up to its junction
with the stream which comes from the west, rising not far from
Leuktra, and, after that, to refer to what is really the main river. From
Platæa itself the course of this stream is plainly traceable in the
plain, running along the western base of the Asopos ridge. The
stream coming from Leuktra is not visible, and it is quite conceivable
that Herodotus never had any definite knowledge of its existence. In
Leake’s time (vide his sketch map) the inhabitants of Kriekouki seem
to have called the stream, A 6, Asopos. It is not so called at the
present day. My own impression is, however, that Herodotus,
although he heard the Platæans speak of A 1 as the Asopos, may in
one passage refer to the stream from Leuktra with a special attribute:
τὸν Ἀσωπὸν τὸν ταύτῃ ῥέοντα (H. ix. 31). A sentence previously, at
the end of Chapter 30, he has a reference to the Asopos without any
qualification, οὗτοι μὲν νὺν ταχθέντες ἐπὶ τῷ Ἀσωπῷ
ἐστρατοπεδἐυοντο, and this reference is undoubtedly to A 1, which is
to him, as other references in his narrative show, the upper Asopos
“ordinarily so called.”
0
H. ix. 31, ad init., πυθόμενοι τοὺς Ἕλληνας εἶναι ἐν Πλαταιῇσι.
Cf. especially the mention of the Asopos and its context in Chapter
40.
2
It will be remarked that Artabazos’ statement on this point is in direct
conflict with that reported by Herodotus to have been made at the
same time by Alexander of Macedon to the Greeks.
3
It appears later (Chap. 46, ad init.) that it was to the Athenian
generals alone that Alexander’s story was in the first instance
imparted. That tends to confirm, what the lie of the ground would
suggest, that the Greek left was nearer the Asopos than the right
wing.
4
This is one of the most important passages in Herodotus’ description
of the battle. It indicates more clearly than has been hitherto
indicated, the position of the Greeks in their second position.
In the first place, if we remember that the Lacedæmonians were
on the Greek right, it will be seen that it forms a very strong
argument in favour of the identification of Gargaphia which has been
adopted. Had it been at Apotripi it would certainly have been near
the Greek centre. It also shows the obliquity of the Greek line with
respect to the course of the Asopos; in other words, that it was, as
might be expected, extended along the Asopos ridge.
5
This is shown still more clearly in the account of the withdrawal from
this position.
6
The three developments of the Greek second position may be
summed up as follows:⁠—
1. The Greek right was near the spring of Gargaphia, not on the
Asopos ridge, while the left was near the Heroön of Androkrates.
2. After a forward movement of the whole line, the right took up
position on the Asopos ridge, while the line extended along the
course of that ridge, until the left was actually on the Asopos.
3. The left, when its position on the plain became untenable, took
to the higher ground of the north extension of the Asopos ridge.
7
It would seem as if this determination were not come to at the
morning council. Their idea at that time appears to have been to
move during the night, in case the enemy did not renew their attack.
As the attack was renewed, the movement was deferred until the
following night.
8
The members of the American school at Athens who excavated
parts of the site of Platæa some years ago were inclined to believe
that at the time of the battle the town stood on the higher or southern
end of the bastion which is now strewn with the traces of the
successive towns which have occupied the site; and that it did not
extend northward to the point where the bastion sinks more or less
abruptly into the plain. They also believed that they discovered the
foundations of the temple of Hera on this north extension of the
bastion. I am disposed to think that their conjecture as to the position
of the contemporary town is correct, though the question is not of
sufficient importance with respect to this particular passage in Greek
history to render it desirable or necessary to quote the mass of
evidence on which the opinion is founded. The position of the temple
of Hera as determined by them agrees with the brief mention of it in
this passage of Herodotus.
9
Herodotus, in words already quoted, says that it was the intention of
the Greeks, on moving to the “Island,” to detach a part of the army to
relieve the attendants who were blocked in the pass. This is certainly
the Dryoskephalæ or the Platæa-Athens pass, probably the latter,
which they were attempting to use as an alternative way, after the
fearful disaster which had befallen the former provision train in the
exit of the Dryoskephalæ pass. Herodotus shows, too, that this relief
was urgently required, since the Greek army was running short of
provisions; for, although the Platæa-Megara pass must have been
open, it is of such a character as to render it impossible that the
commissariat for a force of 100,000 men could be adequately
maintained through its channel. It is therefore in the very highest
degree probable that an attempt, at any rate, was made to carry out
this part of the arrangement between the generals. Now, the Spartan
force on the right of the Greek line would be, in so far as position
was concerned, that portion of the Greek army on which this duty
would naturally devolve. The mission of this force for the relief of the
pass was one of extreme danger and difficulty, and it would be
natural that the service should devolve on that part of the army which
enjoyed the highest military reputation. It was, I venture to think,
while carrying out this movement that the Spartans became involved
in that series of events which led to the last catastrophe in the great
tragedy.
0
Thucydides (i. 20) denies that such a division or regiment existed in
the Spartan army.
Even in the Spartan army indiscipline was apt to make its
appearance without the existence of such a substantial motive as in
the present instance. Cf. the insubordination of the Spartan officers
at the battle of Mantinea in 418 b.c. (Thuc. v. 72).
2
That they never reached the rocky ὑπωρέη is plain from the
incidents of the battle that followed.
3
Of A 4 and A 5.
4
The ὑπωρέη of Herodotus.
5
Cf. the tale H. ix. 58.
6
Δρόμῳ διαβάντας τὸν Ἀσωπὸν (H. ix. 59).
7
It will be seen, when the details of the Athenian retreat come to be
examined, how noticeably this detail accords with the account which
Herodotus gives of that retreat.
8
Some modern commentators have regarded this detail mentioned by
Herodotus as a convincing proof of the Athenian bias in his narrative.
To me it seems eminently natural, after the experience of the
previous days, that Pausanias or any other commander should have
summoned help under the circumstances. I shall, moreover, have
occasion to show that the Athenians did undoubtedly diverge from
their march to the Island in the direction in which the Spartan battle
with the Persians took place.
9
It is clear from Herodotus’ subsequent account of the proceedings of
the Greek centre that this battle took place out of sight of that part of
the army which had retired to Platæa.
0
H. ix. 62: Ἤδη ἐγίνετο ἡ μάχη ἰσχυρὴ παρ’ αὐτὸ τὸ Δημήτριον.
This incidental detail mentioned by Herodotus peculiarly supports
the view that the temple must have stood on the site of the church of
St. Demetrion.
2
This is clearly shown in Herodotus’ narrative. He distinctly speaks of
the Athenians as having at the beginning of the movement “turned
down towards the plain” (H. ix. 56, κάτω τραφθέντες ἐς τὸ πεδίον);
and in a still more remarkable passage he says that, when
Mardonius led his Persians across the Asopos in pursuit of the
Greeks, “he did not see the Athenians, who had turned down
towards the plain, by reason of the (intervening) hills” (H. ix. 59). The
hills mentioned are evidently the northern extension of the Asopos
ridge.
3
A 1 in the map.
4
Ridge 5.
5
Thus far διὰ τῆς ὑπωρέης (H. ix. 69).
6
I.e. ridges 3 and 2; cf. H. ix. 69, διὰ ... τῶν κολωνῶν.
7
Ridge 5.
8
I confess I cannot understand the argument of those who regard
Herodotus’ account of Platæa as being tainted throughout with a
lying Athenian tradition. In so far as the narrative provides evidence
of its source or sources, there is at least as much matter in it which
may be attributed to Spartan as to Athenian origin.
9
The Asopos ridge, the Long ridge, and the Plateau.
0
The treatment meted out to the Æginetans in the narrative of
Platæa, as contrasted with the account which Herodotus gives of
their conduct at Salamis, points to the very various character of the
sources from which he drew his history. This part of the Platæan
narrative is undoubtedly drawn from a tradition highly coloured by
the relations which existed between Athens and Ægina twenty years
after Platæa was fought.
Xen. Anab. iii. 2, 27. The striking words are μὴ τὰ ζεύγη ἡμῶν
στρατηγῇ.
2
I have had occasion to speak of the Thermopylæ narrative under
various aspects in relation to the sources from which it is derived.
To prevent any misconception, I should like to sum up briefly my
conclusions.
(1) The whole “motivation” of the story is derived from a
version of official origin at Sparta.
(2) The incidents of the actual fighting may be derived partly
from a Spartan source, probably of an unofficial character. The
description of some of them, however, rests on information
picked up by Herodotus at Thermopylæ itself from natives of the
region.

Trends in Artificial Intelligence

Uploaded by

Copyright:

Available Formats

Trends in Artificial Intelligence

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Trends in Artificial Intelligence

Uploaded by

Copyright:

Available Formats

Full download test bank at ebook textbookfull.

Trends in Artificial Intelligence

Trends in Artificial Intelligence

Theory and Applications

Subseries of Lecture Notes in Computer Science

Moonis Ali Jun Sasaki (Eds.)

Trends in Artiﬁcial Intelligence

ISSN 0302-9743 ISSN 1611-3349 (electronic)

© Springer Nature Switzerland AG 2020

September 2020 Hamido Fujita

Special Session Chairs

Special Session Organizers

Yu-Chen Lin Feng Chia University, Taiwan

Franz Wotawa TU Graz, Austria

Natural Language Processing

Question Generation Through Transfer Learning . . . . . . . . . . . . . . . . . . . . . 3

KIDER: Knowledge-Infused Document Embedding Representation

Discriminative Features Fusion with BERT for Social Sentiment Analysis . . . 30

Text Sentiment Transfer Methods by Using Sentence Keywords . . . . . . . . . . 36

Robotics and Drones

Path Planning of Mobile Robot Group Based on Neural Networks . . . . . . . . 51

Push Recovery and Active Balancing for Inexpensive Humanoid Robots

Optimal Control Problem of a Differential Drive Robot . . . . . . . . . . . . . . . . 75

Optimal Trajectory Tracking Control for a UAV Based on Linearized

Non-linear Control of Aerial Manipulator Robots Based

Non-linear 3D Visual Control for an Unmanned Aerial Vehicle . . . . . . . . . . 108

Construction and Control Aerial Manipulator Robot . . . . . . . . . . . . . . . . . . 116

Knowledge Based Systems

ConMerge – Arbitration of Constraint-Based Knowledge Bases . . . . . . . . . . 127

A Systematic Model to Model Transformation for Knowledge-Based

Innovative Applications of Intelligent Systems

Mathematical Expression Retrieval in PDFs from the Web Using

Automatic Identification of Account Sharing for Video Streaming Services. . . 162

A Model for Predicting Terrorist Network Lethality and Cohesiveness. . . . . . 174

S2RSCS: An Efficient Scientific Submission Recommendation System

Improved Grey Model by Dragonfly Algorithm for Chinese Tourism

Variable Transformation to a 2 2 Domain Space for Edge

Using Deep Learning Techniques to Detect Rice Diseases from Images

Machine Learning for Water Supply Supervision . . . . . . . . . . . . . . . . . . . . 238

An Enhanced Whale Optimization Algorithm for the Two-Dimensional

A Heuristic Approach to the Three Dimensional Strip Packing Problem

A Heuristic for the Two-Dimensional Irregular Bin Packing Problem

Faster R-CNN Based Fault Detection in Industrial Images . . . . . . . . . . . . . . 280

Estimation of Cable Lines Insulating Materials Resource Using Multistage

User Grouping and Power Allocation in NOMA Systems: A Reinforcement

Deep Learning for QoS-Aware Resource Allocation in Cognitive

Social Network Analysis

Self-understanding Support Tool Using Twitter Sentiment Analysis. . . . . . . . 327

Integrating Crowdsourcing and Active Learning for Classification

Many-to-One Stable Matching for Prediction in Social Networks . . . . . . . . . 345

A Framework for Detecting User’s Psychological Tendencies on Twitter

Automatic Fake News Detection by Exploiting User’s Assessments

Financial Applications and Blockchain

Deep Reinforcement Learning for Foreign Exchange Trading . . . . . . . . . . . . 387

Human-Centred Automated Reasoning for Regulatory Reporting

Security of Blockchain Distributed Ledger Consensus Mechanism

Reinforcement Learning Based Real-Time Pricing in Open Cloud Markets. . . 419

Medical and Health-Related Applications

A New Integer Linear Programming Formulation to the Inverse QSAR/