Computational Linguistics 15th

International Conference of the Pacific

Association for Computational
Linguistics PACLING 2017 Kôiti Hasida
Visit to download the full and correct content document:
https://textbookfull.com/product/computational-linguistics-15th-international-conferenc
e-of-the-pacific-association-for-computational-linguistics-pacling-2017-koiti-hasida/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Computational Linguistics 14th International Conference

of the Pacific Association for Computaitonal
Linguistics PACLING 2015 Bali Indonesia May 19 21 2015
Revised Selected Papers 1st Edition Kôiti Hasida
https://textbookfull.com/product/computational-linguistics-14th-
international-conference-of-the-pacific-association-for-
computaitonal-linguistics-pacling-2015-bali-indonesia-
may-19-21-2015-revised-selected-papers-1st-edition-koiti-ha/

Computational Linguistics 16th International Conference

of the Pacific Association for Computational
Linguistics PACLING 2019 Hanoi Vietnam in Computer and
Information Science 1215 Le-Minh Nguyen (Editor)
https://textbookfull.com/product/computational-linguistics-16th-
international-conference-of-the-pacific-association-for-
computational-linguistics-pacling-2019-hanoi-vietnam-in-computer-
and-information-science-1215-le-minh-nguyen-ed/

Computational Linguistics and Intelligent Text

Processing 18th International Conference CICLing 2017
Budapest Hungary April 17 23 2017 Revised Selected
Papers Part I Alexander Gelbukh
https://textbookfull.com/product/computational-linguistics-and-
intelligent-text-processing-18th-international-conference-
cicling-2017-budapest-hungary-april-17-23-2017-revised-selected-
papers-part-i-alexander-gelbukh/

Computational Linguistics and Intelligent Text

Processing 18th International Conference CICLing 2017
Budapest Hungary April 17 23 2017 Revised Selected
Papers Part II Alexander Gelbukh
https://textbookfull.com/product/computational-linguistics-and-
intelligent-text-processing-18th-international-conference-
cicling-2017-budapest-hungary-april-17-23-2017-revised-selected-
Chinese Computational Linguistics 18th China National
Conference CCL 2019 Kunming China October 18 20 2019
Proceedings Maosong Sun

https://textbookfull.com/product/chinese-computational-
linguistics-18th-china-national-conference-ccl-2019-kunming-
china-october-18-20-2019-proceedings-maosong-sun/

Recent Advances in Computational Engineering

Proceedings of the 4th International Conference on
Computational Engineering ICCE 2017 in Darmstadt
Michael Schäfer
https://textbookfull.com/product/recent-advances-in-
computational-engineering-proceedings-of-the-4th-international-
conference-on-computational-engineering-icce-2017-in-darmstadt-
michael-schafer/

Chinese Computational Linguistics and Natural Language

Processing Based on Naturally Annotated Big Data 16th
China National Conference CCL 2017 and 5th
International Symposium NLP NABD 2017 Nanjing China
October 13 15 2017 Proceedings 1st Edition Maosong Sun
https://textbookfull.com/product/chinese-computational-
linguistics-and-natural-language-processing-based-on-naturally-
annotated-big-data-16th-china-national-conference-
ccl-2017-and-5th-international-symposium-nlp-nabd-2017-nanjing-
ch/
Computational Linguistics and Intelligent Text
Processing 17th International Conference CICLing 2016
Konya Turkey April 3 9 2016 Revised Selected Papers
Part I Alexander Gelbukh
https://textbookfull.com/product/computational-linguistics-and-
intelligent-text-processing-17th-international-conference-
cicling-2016-konya-turkey-april-3-9-2016-revised-selected-papers-
part-i-alexander-gelbukh/

Computational Linguistics and Intelligent Text

Processing 17th International Conference CICLing 2016
Konya Turkey April 3 9 2016 Revised Selected Papers
Part II Alexander Gelbukh
https://textbookfull.com/product/computational-linguistics-and-
intelligent-text-processing-17th-international-conference-
cicling-2016-konya-turkey-april-3-9-2016-revised-selected-papers-
Kôiti Hasida
Win Pa Pa (Eds.)

Communications in Computer and Information Science 781

Computational
Linguistics
15th International Conference of the Pacific Association
for Computational Linguistics, PACLING 2017
Yangon, Myanmar, August 16–18, 2017
Revised Selected Papers

123
Communications
in Computer and Information Science 781
Commenced Publication in 2007
Founding and Former Series Editors:
Alfredo Cuzzocrea, Xiaoyong Du, Orhun Kara, Ting Liu, Dominik Ślęzak,
and Xiaokang Yang

Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Phoebe Chen
La Trobe University, Melbourne, Australia
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
Junsong Yuan
Nanyang Technological University, Singapore, Singapore
Lizhu Zhou
Tsinghua University, Beijing, China
More information about this series at http://www.springer.com/series/7899
Kôiti Hasida Win Pa Pa (Eds.)
•

Computational
Linguistics
15th International Conference of the Pacific Association
for Computational Linguistics, PACLING 2017
Yangon, Myanmar, August 16–18, 2017
Revised Selected Papers

123
Editors
Kôiti Hasida Win Pa Pa
Graduate School of Information Science Natural Language Processing Lab
and Technology University of Computer Studies, Yangon
The University of Tokyo Yangon
Tokyo Myanmar
Japan

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-981-10-8437-9 ISBN 978-981-10-8438-6 (eBook)
https://doi.org/10.1007/978-981-10-8438-6

Library of Congress Control Number: 2018935886

© Springer Nature Singapore Pte Ltd. 2018

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, express 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.

Printed on acid-free paper

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
part of Springer Nature
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
 Preface

This volume is a compilation of selected papers from PACLING 2017, which was the
15th in the series of conferences held since 1989. Most of these events were held in
Australia, Japan, and Canada, but recently we have gathered in developing countries:
Malaysia in 2011, Indonesia in 2015, and Myanmar this time.
This shift coincides with the prospect that these latter countries, among others, have
bigger near-future potentials in realizing smart societies powered by the circulation of
rich and abundant data. For instance, India is building a presence-less, paperless, and
cashless service infrastructure consisting of a nationwide person authentication system,
open-API private and public services, a national PDS (personal data store) for each
individual to coordinate these services while utilizing their personal data. Cambodia
also aims at a cashless society based on a still lighter infrastructure.
It is very likely that most other Asian countries share similar visions of near-future
societies powered by data circulation. Such restructurings are not only inexpensive
enough for developing countries, but also much easier in those countries than in
advanced countries facing stronger opposition by many vested interests. I hence feel
this year's PACLING in Myanmar, with a much higher literacy rate than in India and
Cambodia, is exciting.
I hope the conference and the proceedings contribute to the construction of smart
societies, as technologies to deal with language could address essential parts of the data
circulation.

December 2017 Kôiti Hasida

 Organization

Honorary Chair
Kôiti Hasida The University of Tokyo, Japan

Local Organizing Chair

Mie Mie Thet Thwin University of Computer Studies, Yangon, Myanmar

Local Organizing Committee

Nang Saing Moon Kham University of Computer Studies, Yangon, Myanmar
Win Pa Pa University of Computer Studies, Yangon, Myanmar

Program Committee
Kenji Araki Hokkaido University, Japan
Eiji Aramaki NAIST, Japan
Normaziah Abdul Aziz International Islamic University, Malaysia
Tetsuro Chino Toshiba Corporation, Japan
Khalid Choukri ELRA/ELDA, France
Koji Dosaka Akita Prefectural University, Japan
Alexander Gelbukh Instituto Politécnico Nacional (IPN), Mexico
Li Haizhou National University of Singapore, Singapore
Yoshihiko Hayashi Waseda University, Japan
Tin Myat Htwe KyaingTon Computer University, Myanmar
Bowen Hui University of British Columbia, Canada
Kentaro Inui Tohoku University, Japan
Kai Ishikawa NEC Corporation, Japan
Hiroyuki Kameda Tokyo University of Technology, Japan
Vlado Kesel Dalhousie University, Canada
Satoshi Kinoshita JAPIO, Japan
Kiyoshi Kogure Kanazawa Institute of Technology, Japan
Qin Lu The Hong Kong Polytechnic University, SAR China
Joseph Mariani LIMSI-CNRS, France
Robert Mercer The University of Western Ontario, Canada
Diego Mollá-Aliod Macquarie University, Australia
Hiromi Nakaiwa Nagoya University, Japan
Tin Htar New Magway Computer University, Myanmar
Fumihito Nishino Fujitsu Laboratories, Japan
Win Pa Pa University of Computer Studies, Yangon, Myanmar
VIII Organization

Hamman Riza Agency for the Assessment and Application

of Technology (BPPT), Indonesia
Hiroaki Saito Keio University, Japan
Kazutaka Shimada Kyushu Institute of Technology, Japan
Akira Shimazu Japan Advanced Institute of Science and Technology,
Japan
Kiyoaki Shirai Japan Advanced Institute of Science and Technology,
Japan
Khin Mar Soe University of Computer Studies, Yangon, Myanmar
Virach Sornlertlamvanich Thammasat University, Thailand
Thepchai Supnithi NECTEC, Thailand
Hisami Suzuki Microsoft, USA
Masami Suzuki KDDI Research, Japan
Kumiko Tanaka The University of Tokyo, Japan
Thanaruk Theeramunkong Thammasat University, Thailand
Aye Thida University of Computer Studies, Mandalay, Myanmar
Takenobu Tokunaga Tokyo Institute of Technology, Japan
Mutsuko Tomokiyo Université Grenoble Alpes, France
Yang Xiang University of Guelph, Canada
Yuzana Sittwe Computer University, Myanmar
Ingrid Zukerman Monash University, Australia

Ministry of Education,
Myanmar

University of Computer
Studies, Yangon,
Myanmar
 Contents

Semantics and Semantic Analysis

Detecting Earthquake Survivors with Serious Mental Affliction. . . . . . . . . . . 3

Tatsuya Aoki, Katsumasa Yoshikawa, Tetsuya Nasukawa,
Hiroya Takamura, and Manabu Okumura

A Deep Neural Architecture for Sentence-Level Sentiment Classification

in Twitter Social Networking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Huy Nguyen and Minh-Le Nguyen

Learning Word Embeddings for Aspect-Based Sentiment Analysis . . . . . . . . 28

Duc-Hong Pham, Anh-Cuong Le, and Thi-Kim-Chung Le

BolLy: Annotation of Sentiment Polarity in Bollywood Lyrics Dataset . . . . . 41

G. Drushti Apoorva and Radhika Mamidi

Frame-Based Semantic Patterns for Relation Extraction . . . . . . . . . . . . . . . . 51

Angrosh Mandya, Danushka Bollegala, Frans Coenen,
and Katie Atkinson

Semantic Refinement GRU-Based Neural Language Generation for Spoken

Dialogue Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Van-Khanh Tran and Le-Minh Nguyen

Discovering Representative Space for Relational Similarity Measurement . . . . 76

Huda Hakami, Angrosh Mandya, and Danushka Bollegala

Norms of Valence and Arousal for 2,076 Chinese 4-Character Words . . . . . . 88

Pingping Liu, Minglei Li, Qin Lu, and Buxin Han

Statistical Machine Translation

Integrating Specialized Bilingual Lexicons of Multiword Expressions

for Domain Adaptation in Statistical Machine Translation . . . . . . . . . . . . . . 101
Nasredine Semmar and Meriama Laib

Logical Parsing from Natural Language Based on a Neural

Translation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Liang Li, Yifan Liu, Zengchang Qin, Pengyu Li, and Tao Wan
X Contents

Phrase-Level Grouping for Lexical Gap Resolution

in Korean-Vietnamese SMT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Seung Woo Cho, Eui-Hyeon Lee, and Jong-Hyeok Lee

Enhancing Pivot Translation Using Grammatical

and Morphological Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Hai-Long Trieu and Le-Minh Nguyen

Corpora and Corpus-Based Language Processing

Information-Structure Annotation of the “Balanced Corpus

of Contemporary Written Japanese”. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Takuya Miyauchi, Masayuki Asahara, Natsuko Nakagawa,
and Sachi Kato

Syntax and Syntactic Analysis

Khmer POS Tagging Using Conditional Random Fields . . . . . . . . . . . . . . . 169

Sokunsatya Sangvat and Charnyote Pluempitiwiriyawej

Statistical Khmer Name Romanization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

Chenchen Ding, Vichet Chea, Masao Utiyama, Eiichiro Sumita,
Sethserey Sam, and Sopheap Seng

Burmese (Myanmar) Name Romanization: A Sub-syllabic Segmentation

Scheme for Statistical Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Chenchen Ding, Win Pa Pa, Masao Utiyama, and Eiichiro Sumita

Document Classification

Domain Adaptation for Document Classification by Alternately Using

Semi-supervised Learning and Feature Weighted Learning . . . . . . . . . . . . . . 205
Hiroyuki Shinnou, Kanako Komiya, and Minoru Sasaki

Information Extraction and Text Mining

End-to-End Recurrent Neural Network Models for Vietnamese

Named Entity Recognition: Word-Level Vs. Character-Level . . . . . . . . . . . . 219
Thai-Hoang Pham and Phuong Le-Hong

Nested Named Entity Recognition Using Multilayer Recurrent

Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Truong-Son Nguyen and Le-Minh Nguyen
 Contents XI

Text Summarization

Deletion-Based Sentence Compression Using Bi-enc-dec LSTM . . . . . . . . . . 249

Dac-Viet Lai, Nguyen Truong Son, and Nguyen Le Minh

Text and Message Understanding

Myanmar Number Normalization for Text-to-Speech . . . . . . . . . . . . . . . . . . 263

Aye Mya Hlaing, Win Pa Pa, and Ye Kyaw Thu

Expect the Unexpected: Harnessing Sentence Completion

for Sarcasm Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Aditya Joshi, Samarth Agrawal, Pushpak Bhattacharyya,
and Mark J. Carman

Detecting Computer-Generated Text Using Fluency and Noise Features . . . . . 288

Hoang-Quoc Nguyen-Son and Isao Echizen

Automatic Speech Recognition

Speaker Adaptation on Myanmar Spontaneous Speech Recognition . . . . . . . . 303

Hay Mar Soe Naing and Win Pa Pa

Exploring the Effect of Tones for Myanmar Language Speech Recognition

Using Convolutional Neural Network (CNN) . . . . . . . . . . . . . . . . . . . . . . . 314
Aye Nyein Mon, Win Pa Pa, and Ye Kyaw Thu

Spoken Language and Dialogue

Listenability Measurement Based on Learners’ Transcription Performance . . . 329

Katsunori Kotani and Takehiko Yoshimi

Speech Pathology

A Robust Algorithm for Pathological-Speech Correction . . . . . . . . . . . . . . . 341

Naim Terbeh and Mounir Zrigui

Speech Analysis

Identification of Pronunciation Defects in Spoken Arabic Language . . . . . . . 355

Naim Terbeh and Mounir Zrigui

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

Semantics and Semantic Analysis
 Detecting Earthquake Survivors
with Serious Mental Affliction

Tatsuya Aoki1(B) , Katsumasa Yoshikawa2 , Tetsuya Nasukawa2 ,

Hiroya Takamura1 , and Manabu Okumura1
1
Laboratory for Future Interdisciplinary Research of Science and Technology,
Tokyo Institute of Technology, Yokohama, Japan
{aoki,takamura,oku}@lr.pi.titech.ac.jp
2
IBM Research - Tokyo, IBM Japan, Ltd., Tokyo, Japan
{KATSUY,NASUKAWA}@jp.ibm.com

Abstract. The 2011 Great East Japan Earthquake and 2016 Kuma-
moto earthquakes had a great impact on numerous people all over the
world. In this paper, we focus on social media and the mental health
of 2016 Kumamoto earthquake survivors. We first focus on the users
who had experienced an earthquake and track their sentiments before
and after the disaster using Twitter as a sensor. Consequently, we found
that their emotional polarities switch from nervous during earthquakes
and return to normal after huge earthquakes. However, we also found
that some people did not go back to normal even after huge earthquakes
subside. Against this background, we attempted to identify survivors who
are suffering from serious mental distress concerning earthquakes. Our
experimental results suggest that, besides the frequency of words related
to earthquakes, the deviation in sentiment and lexical factors during the
earthquake represent the mental conditions of Twitter users. We believe
that the findings of this study will contribute to early mental health care
for people suffering the aftereffects of a huge disaster.

1 Introduction
The Kumamoto earthquakes occurred on April 14th, 2016. Earthquakes of mag-
nitude 7.0 occurred twice, and these huge shocks caused severe damage. As a
result, these earthquakes damaged over 185,000 houses and 4,600 buildings, killed
at least 50 people, and made approximately 180,000 people evacuees. Moreover,
over 3,700 aftershocks occurred from April to June. The long-term aftershocks
are a reason for the over 100 disaster-related deaths, including deaths indirectly
related to the disaster such as death due to disaster-related stress. Additionally,
because of the many aftershocks, some survivors suffered repeated post-quake
trauma resulting in post-traumatic stress disorder.
In such an unusual condition, a microblog is a helpful tool to inform others
of the current situation immediately. Many Twitter users tweeted about their
safety or the emergency situation during the earthquakes. After the earthquakes,
the survivors tended to use microblogs to contact their friends or people who
c Springer Nature Singapore Pte Ltd. 2018
K. Hasida and W. P. Pa (Eds.): PACLING 2017, CCIS 781, pp. 3–14, 2018.
https://doi.org/10.1007/978-981-10-8438-6_1
4 T. Aoki et al.

were in the same situation in order to share disaster-related information such as

evacuation information and also to release anxiety by sending messages to each
other. For earthquake survivors, information such as the status of the water sup-
ply is important to let people know their circumstances in a huge disaster. This
kind of information is also important support because we are able to discover
help messages based on the emergency messages from social media.
While many researchers have focused on retrieval of information regarding
natural disasters [12,14,18], few have focused on the survivors’ behaviors and
their mental condition. Posts that exhibit signals of mental illness caused by
unusual conditions increase during a disaster. The characteristics indicate that
there are people in need of mental health care because of disaster-related mental
aftereffects.
In this paper, we aim to identify people suffering severe mental stress due to
earthquake. We also show the importance of social media to mental health by
analyzing posts on Twitter of Kumamoto earthquake survivors. For constructing
the dataset, we annotated the posts of Kumamoto survivors with earthquake-
related stress levels judged by the content of their posts. Although there are
people suffering from mental aftereffects of earthquakes, there has not been prior
work on social media analysis that mentions mental health care for those people.
It would be worthwhile to try to catch signals of mental illness from social media
because we are not able to identify who needs mental health care after a large-
scale disaster. We believe that this study will contribute to early mental health
care for people suffering the aftereffects of a huge disaster.
Our main contributions in this study are as follows:
– We attempt to examine the relation between a huge earthquake and shifts in
mental conditions by leveraging social media.
– We conduct a large-scale sentiment analysis using time series social media
data. We track the sentiments of earthquake survivors before and after a
huge disaster.
– We show the possibility that we can detect people who are suffering from seri-
ous mental stress concerning earthquakes on the basis of sentiment deviation
and language use on social media.

2 Related Work
There is a wide range of research that uses microblogs as a sensor for detecting
events [16], predicting geolocation [2], mining public opinion [13] and catching
tendency [1,10].
Vieweg et al. [19] suggested that microblogged information plays a role of
a situational awareness observer when an emergency event occurs. There are
also several research efforts that leverage microblogged information during nat-
ural disaster events for facilitating support [18]. Okazaki et al. [14] collected
misinformation automatically in order to prevent false information from spread-
ing through social media. Neubig et al. [12] studied ways of extracting safety
information about people during the 2011 East Japan Earthquake from Twitter.
 Detecting Earthquake Survivors with Serious Mental Affliction 5

Table 1. Dataset for sentiment analysis

Period March 1st, 2016–June 6th, 2016

#User 2,629
#Tweet 2,729,245

Varga et al. [18] built an information retrieval system to extract problem reports
and aid messages from Twitter during natural disasters.
Recent studies have suggested that Twitter is a plentiful source of public
health. Twitter allows us to access a large amount of information about pub-
lic health including mental disorders. Previous studies connecting social media
to mental health use knowledge from social data for predicting depression [4],
measuring stress level [11], and exploring mental health and language [5,7]
Nevertheless, there have been few investigations of the relation between real
events and mental condition shifts which can be seen in social media. Choudhury
et al. [3] studied the behavioral changes of new mothers and predicted significant
postnatal behavioral changes based on prenatal information gathered on Twitter.
Kumar et al. [9] focused on celebrity suicides that caused suicides of their fans
and analyzed posting frequencies and content on Reddit after celebrity suicides
by using a topic model.

3 Time Series Sentiment Analysis

As a preliminary experiment, we examined the sentiment of survivors before
and after the Kumamoto earthquakes by constructing a sentiment model that
measures the sentiment of users from their posts to a microblog. We used Twitter
as a dataset and collected posts between March 1st, 2016 and June 6th, 2016. To
focus on the Kumamoto earthquakes survivors, we analyzed only Twitter users
whose location in the user profile or tweet location was Kumamoto. Table 1
summarizes our dataset.

3.1 Sentiment Model

First, we define the sentiment model and sentiment score. The sentiment model
measures the user’s sentiment level. The output of the sentiment model is a senti-
ment score indicating the sentiment levels of users. We built the sentiment model
by using tweets as input. Let D be a set of tweets. Dpositive and Dnegative ⊆ D
are the sets of positive tweets and negative tweets respectively, which are subsets
of D. Dpolarity is the set of tweets with positive/negative sentiment, and defined
as Dpolarity = Dpositive ∪ Dnegative . The sentiments of tweets are determined
by the sentiment analyzer of IBMWatson
R Explorer Advanced Edition Ana-
lytical Components V11.0 (WEX)1 . However, only a few tweets were provided
1
IBMWatson
R Explorer Advanced Edition Analytical Components V11.0 is a trade-
mark in the United States and/or other countries of International Business Machines
Corporation.
6 T. Aoki et al.

Fig. 1. Time Series Sentiment Score

with sentiments determined with the WEX sentiment analyzer because of errors
regarding the dependency analysis and morphological analysis for short texts
written in colloquial style. Thus, we used pattern matching with the Japanese
Sentiment Dictionary [8,17] in addition to the sentiment analyzer. If the results
of the two methods contradict each other, we used both estimation results as the
sentiment of the tweets. We thus permitted both positive and negative sentiments
to be allocated to a tweet. The sentiment scoring function s(D) is defined as:

s(D) = g1 (Dpositive , Dnegative ) × g2 (D, Dpolarity ),

|Dpositive | − |Dnegative |
g1 (Dpositive , Dnegative ) = ,
|Dpositive | + |Dnegative |
|Dpolarity |
g2 (D, Dpolarity ) = .
|D|

The sentiment scoring function s(D) consists of two factors, where

g1 (Dpositive , Dnegative ) returns indicates the inclination of tweets toward pos-
itive sentiment and g2 (D, Dpolarity ) returns the proportion of sentiment tweets.

3.2 Time Series Sentiment Score

For the dataset shown in Table 1, we demonstrate how the sentiment score
changes day by day. We excluded tweets containing earthquake-related words2 ,
earthquake, aftershock, immediate report, shelter, shaking and goods, from
Dnegative in order to prevent the sentiment score from being biased toward neg-
ative due to those words; we would like to focus on the users’ mental conditions.
2
We extracted expressions which occur with high frequency and correlation by using
WEX and defined them as earthquake-related words: (earthquake), (after-
shock), (immediate report), (shelter), (shaking) and (goods).
 Detecting Earthquake Survivors with Serious Mental Affliction 7

Figure 1 shows the average daily sentiment score of the users in our dataset.
As shown in Fig. 1, the sentiment score suddenly drops on April 14th, i.e., when
the Kumamoto Earthquakes started, and then returns to the same tendency
before the drop. The graph marks the lowest score (0.048pt) on April 16th,
when the main shock of the Kumamoto Earthquakes occurred.

Table 2. Average sentiment scores by period

Period Date AVE. score #Tweet

Before disaster Mar. 1–Apr.13 0.134 1,016,068
During disaster Apr. 13–Apr. 23 0.103 366,154
After disaster Apr. 24–Jun. 6 0.131 1,347,023

Table 3. Average sentiment scores on different dataset

Group AVE. score #Tweet

Control 0.119 8,043
Schizophrenia 0.049 36,417
ADHD 0.042 29,454
Depression 0.029 38,782
PTSD 0.012 43,578
Borderline 0.004 29,648
Bipolar −0.000 20,039

3.3 Comparison of Sentiment Score

To evaluate the sentiment scores before and after the earthquake, we split the
dataset into three parts by date. We defined the 44 days before the first earth-
quake (April 14) as the before-disaster period, the 10 days after the first earth-
quake occurred as the during-disaster period, and the 44 days after the during-
disaster period as the after-disaster period.
Table 2 summarizes the average sentiment scores within the whole period. As
shown in Table 2, the score for the during-disaster period is 0.103pt and approxi-
mately 0.03pt lower than the scores of the before- and after-disaster periods. On
the other hand, the scores are at the same level in the before- and after-disaster
periods, which indicates that the sentiment scores of users returned to neutral
after the huge earthquakes.
To investigate the relation between the mental health condition and the sen-
timent score, we built a new dataset to compare the scores on different groups:
a control group and a mental-health-aware group. As the control group, we ran-
domly picked users from Twitter and collected their most recent 150 tweets on
October, 2016. As the mental health-aware group, we selected Twitter users who
8 T. Aoki et al.

Table 4. Example tweets of survivors after huge earthquakes

Table 5. Dataset for serious survivor detection

Period March 1st, 2016–June 6th, 2016

#User 289
#positive 29
#negative 260
#Tweet 366,154

had written the name of an illness in their Twitter profile and collected their most
recent 150 tweets in October, 2016. We used the self diagnosis profile matching
method that finds the users whose twitter profiles contain a name of a mental
disorder. We defined the mental-health-aware group to be these users. Following
the research of Coppersmith et al. [6], we focused on six mental health conditions:
schizophrenia, attention deficit hyperactivity disorder (ADHD), depression, post-
traumatic stress disorder (PTSD), borderline personality disorder (Borderline),
and bipolar disorder (Bipolar). Next, we compared the sentiment score of the
control group and those of other groups. For calculating the sentiment score, we
regarded the tweets of the group as one large set of tweets.
Table 3 summarizes the average sentiment score of the control group and the
six mental-health-aware groups. Compared with the score of the control group,
the scores of the mental-health-aware groups tended to be lower. The average
score of mental-health-aware groups was 0.023pt, close to the lowest score in
Fig. 1 (0.048pt), i.e., the sentiment score on April 16th when the main shock of
the Kumamoto Earthquakes occurred.

4 Seriously Affected Survivor Detection

4.1 Problem Setting
Our dataset included tweets from people suffering from aftereffects of the earth-
quake. Table 4 shows example tweets of survivors. Their mental condition might
be negatively affected by their experiences in the earthquake. Such people would
 Detecting Earthquake Survivors with Serious Mental Affliction 9

require early mental care. Accordingly, we tried to detect users who exhibited
signals of mental illnesses caused by the earthquake. To tackle this detection
task, we built a new dataset that was a subset of the dataset used in Sect. 3 and
randomly picked 300 users. We annotated each of the 300 users with one of two
class labels: a user who does not care about disaster and a user who is severely
affected by the disaster, on the basis of their mental health conditions deter-
mined from their tweets3 . Since examining all tweets for each user is too costly
for annotation purposes, we only examined the tweets containing earthquake-
related words. We defined positive users as those whom both annotators labeled
“affected” and negative users as otherwise. Table 5 shows the statistics of this
dataset.

4.2 Methodology
Our goal for this task is to detect 29 positive users from 289 users. Since a
large labeled dataset for this task is not available, we developed an unsupervised
method, which relies on a final score indicating poor mental health of each user.
The final score was calculated as the product of three factors described below.

Lexical Factor. It is supposed that users in poor mental health would have
particular word usage characteristics. The lexical factor was calculated as the
sum of three components: proportion of earthquake-related words, proportion of
words associated with sleep disorders, and linguistic deviation of tweets.

– Proportion of Earthquake-related Words

Users concerned with the earthquake might be more likely to be mentally
disordered. Thus, as a basic factor, we used the proportion of earthquake-
related words |Deq−words |/|D| , where Deq−words is the set of tweets that
contain earthquake-related words and D is the set of all tweets of the user.
When the proportion of earthquake-related words is used together with other
factors, we used its following normalized version for the purpose of keeping a
good balance with others:
|Deq−words | |Deq−words |
× ,
|Dalleq−words | |D|

where Dalleq−words is the set of tweets that contain earthquake-related words

in the entire dataset.
– Proportion of Words Associated with Sleep-Disorder
Users in poor mental health often suffer from sleep disorders. We thus used
|Dsleep−disorder |/|D| , where Dsleep−disorder is the set of tweets that contain
3
The actual annotation was a three class label: a user who does not care about the
disaster, a user who is slightly concerned with the disaster, and user who is severely
affected by the disaster. This annotation was conducted by two human annotators,
and Cohen’s kappa for that is 0.75. By merging the first two classes into one class,
we created an experimental dataset with two classes.
10 T. Aoki et al.

the words associated with sleep disorders. To obtain the tweets expressing
sleep disorders, we collected expressions related to sleep disorders and added
them to the dictionary of WEX.
– Emotion Bias
Users in poor mental health would more likely express anxiety and fear rather
than optimistic emotions. We therefore used the following emotion deviation
factor:
|Danxiety | + |Df ear | − |Doptimistic |
,
|D|
where Danxiety , Df ear , and Doptimistic are determined by a classifier for lin-
guistic categories based on Linguistic Inquiry and Word Count [15].

User’s Vulnerability. We supposed that vulnerable users would take negative

events more seriously. To implement this idea, we used Personality Insights4
to measure the vulnerability of each user. Personality Insights is a personality
estimation service, which analyzes user personality in terms of psychology by
feeding tweets into the system. We used its output to calculate the following
factor:
V ulnerability(Dduring ) + V ulnerability(Daf ter )
−V ulnerability(Dbef ore ),
where Dbef ore , Dduring , and Daf ter are respectively the sets of tweets posted in
the periods of before, during, and after the disaster , as shown in Table 2.

Sentiment Deviation. We focused on the difference between the sentiment

score of a user and the average score of all users. We additionally calculated
earthquake-related sentiment scores with the same method as in Fig. 1. With
the same method as in Fig. 1, we recalculated the sentiment scores for only
tweets containing earthquake-related words; we called it the “earthquake-related
sentiment score”. When calculating the sentiment scores of day k, we used the
set of tweets posted from k to k+10 in order to alleviate any tweet sparseness of a
user. We calculated two kinds of daily sentiment score. Then, we used these daily
sentiment scores to calculate two deviation values. This factor is expressed as
1 1 
n m
(xi − xi )2 + (yj − yj )2 ,
n i=1 m j=1

where xi is the sentiment score of the i-th day when the user posted a tweet
and xi is the average sentiment score of day i in the entire dataset. n is the
number of days when the user posted a tweet. Similarly, yj is the earthquake-
related sentiment score of the j-th day when the user posted a tweet containing
earthquake-related words and yj is the average earthquake-related sentiment
score of day j in the entire dataset. m is the number of days when the user
posted a tweet that contains earthquake-related words.
4
https://personality-insights-livedemo.mybluemix.net.
 Detecting Earthquake Survivors with Serious Mental Affliction 11

4.3 Experiment
We evaluated our model using two measures. One was Recall@30 (R@30), which
examined how many positive users are included in the top 30 users who have
the highest final score. The other was average precision (AP), which investigated
how much higher positive users are ranked compared with other users. Note that
in the following ALL denotes the proposed method in which all factors were used.

Table 6. Results of Recall@30 and average precision

Feature #detect user R@30 AP

Proportion of Earthquake-related Words 10 0.35 0.32
+Vulnerability 11 0.38 0.37
+Vulnerability+Lexical Factor 14 0.48 0.41
+Vulnerability+Lexical Factor+Sentiment Deviation 17 0.59 0.45

Table 7. Ablation test of proposed model

Feature #detect user R@30 AP

ALL (Proposed) 17 0.59 0.45
w/o Proportion of EQ Words 16 0.55 0.45
w/o Vulnerability 16 0.55 0.46
w/o Lexical Factor 14 0.48 0.43
w/o Sentiment Deviation 14 0.48 0.41

Quantitative Evaluation. Table 6 shows that our model outperformed the

other methods, achieving 0.59 pt in Recall@30 and 0.45 pt in AP.
Table 7 summarizes the results of the ablation test for the proposed model
to determine which feature is effective for this task. First, we discuss the results
of Recall@30. As presented in Table 7, the models without the proportion of
earthquake-related words and user vulnerability features from the proposed
model each performed 0.04pt worse than the proposed model. In addition, the
models without sentiment deviation and lexical factor performed 0.11pt worse.
Thus, the lexical factor and sentiment deviation have much more impact than
the other two in our task.
Almost the same tendency as in Recall@30 can be seen on the results of
average precision (AP). For the AP evaluation, the exclusion of lexical factors
or sentiment deviations severely hurt performance. The only difference between
Recall@30 and AP is that the model without the user’s vulnerability achieved
the highest score at 0.46pt. However, we considered that the tiny improvement
(0.01pt margin) is within the margin of error. Thus, these results indicate that
the sentiment deviation and lexical factor are effective features for detecting
positive users.
12 T. Aoki et al.

Table 8. Example tweets of positive user ranked seventh by the proposed model

Table 9. Example tweets of ordinary user ranked second by the proposed model

Qualitative Evaluation. Besides the quantitative evaluation, we focused on

users labeled as “mental-health-aware” and checked the content of their tweets.
Table 8 shows some posts by a user ranked by the proposed model at 7 out of
281. As shown in Table 8, the user’s posts describe his/her sleep disorder, and it
can be seen that the user has negative feelings about the earthquakes. Moreover,
as can be seen in the last sentence of Table 8 indicating that the earthquakes
were a traumatic experience for him/her, this user was mentally affected by the
earthquakes.

Error Analysis and Discussions. This section describes the error analysis
and discusses the results. First, we report the negative users whom our model
incorrectly detected as positive. Table 9 shows a user erroneously ranked in sec-
ond place. Although posts related to earthquakes by this user increased after the
huge earthquakes, there seemed to be no signal of mental illness.
Next, we report the positive users who were not detected by our model.
Table 10 shows some examples of tweets of an undetected user who was the
lowest-ranked positive user by the proposed model. Although the user is ranked

Table 10. Example tweets of positive user ranked 268th by the proposed model
 Detecting Earthquake Survivors with Serious Mental Affliction 13

268 by our model, it is highly probable that the user was mentally affected by
the earthquakes. This is because our model was not able to capture the implicit
expressions of the user. In order to capture the emotions of users from tweets, we
should consider errors in preprocessing such as by using morphological analysis
to process tweets that are usually written in a colloquial style.
Finally, we discuss the limitations of our work. One of the limitations is that
our model does not take into account the temporal aspects of posting activities.
For example, if the posting schedule suddenly changed after a huge disaster, it
suggests there was some change in the user’s life. Thus, if a user suddenly stops
posting tweets, we would have to be careful with that user. Leveraging such
temporal aspects would improve the ability of the method.

5 Conclusions

We aimed to detect the survivors who may suffer from mental illness due to
the experience of a disaster. Our experimental results suggest that sentiment
deviation and lexical factors including proportion of words associated with sleep
disorders and emotion bias are effective features for detecting users who likely
suffer from mental illness.
In the future, we will apply our method to mental health checks at school
and companies by leveraging external information such as daily reports.

Acknowledgements. We are grateful to Koichi Kamijoh, Hiroshi Kanayama,

Masayasu Muraoka, and the members of the IBM Research - Tokyo text mining team
for their helpful discussions.

References
1. Aramaki, E., Maskawa, S., Morita, M.: Twitter catches the flu: detecting influenza
epidemics using Twitter. In: Proceedings of EMNLP 2011, pp. 1568–1576 (2011)
2. Backstrom, L., Sun, E., Marlow, C.: Find me if you can: improving geographical
prediction with social and spatial proximity. In: Proceedings of WWW 2010, pp.
61–70 (2010)
3. Choudhury, M.D., Counts, S., Horvitz, E.: Predicting postpartum changes in emo-
tion and behavior via social media. In: Proceedings of CHI 2013, pp. 3267–3276
(2013)
4. Choudhury, M.D., Gamon, M., Counts, S., Horvitz, E.: Predicting depression via
social media. In: Proceedings of ICWSM 2013 (2013)
5. Coppersmith, G., Dredze, M., Harman, C.: Quantifying mental health signals in
Twitter. In: Proceedings of Workshop on Computational Linguistics and Clinical
Psychology, pp. 51–60 (2014)
6. Coppersmith, G., Dredze, M., Harman, C., Hollingshead, K.: From ADHD to SAD:
analyzing the language of mental health on Twitter through self-reported diag-
noses. In: Proceedings of Workshop on Computational Linguistics and Clinical
Psychology, pp. 1–10 (2015)
14 T. Aoki et al.

7. Gkotsis, G., Oellrich, A., Hubbard, T., Dobson, R., Liakata, M., Velupillai, S.,
Dutta, R.: The language of mental health problems in social media. In: Proceedings
of Workshop on Computational Linguistics and Clinical Psychology, pp. 63–73
(2016)
8. Higashiyama, M., Inui, K., Matsumoto, Y.: Learning sentiment of nouns from
selectional preferences of verbs and adjectives. In: Proceedings of Annual Meeting
of the Association for Natural Language Processing (in Japanese), pp. 584–587
(2008)
9. Kumar, M., Dredze, M., Coppersmith, G., Choudhury, M.D.: Detecting changes
in suicide content manifested in social media following celebrity suicides. In: Pro-
ceedings of HT 2015, pp. 85–94 (2015)
10. Lamb, A., Paul, M.J., Dredze, M.: Separating fact from fear: tracking flu infections
on Twitter. In: Proceedings of NAACL:HLT 2013, pp. 789–795 (2013)
11. Lin, H., Jia, J., Nie, L., Shen, G., Chua, T.: What does social media say about
your stress? In: Proceedings of IJCAI 2016, pp. 3775–3781 (2016)
12. Neubig, G., Matsubayashi, Y., Hagiwara, M., Murakami, K.: Safety information
mining-what can NLP do in a disaster-. In: Proceedings of IJCNLP 2011, pp.
965–973 (2011)
13. O’Connor, B., Balasubramanyan, R., Routledge, B.R., Smith, N.A.: From tweets
to polls: linking text sentiment to public opinion time series. In: Proceedings of
ICWSM 2010, pp. 23–26 (2010)
14. Okazaki, N., Nabeshima, K., Watanabe, K., Mizuno, J., Inui, K.: Extracting and
aggregating false information from microblogs. In: Proceedings of Workshop on
Language Processing and Crisis Information, pp. 36–43 (2013)
15. Pennebaker, J.W., Francis, M.E., Booth, R.J.: Linguistic inquiry and word count:
LIWC 2001. Lawrence Erlbaum Associates (2001)
16. Sakaki, T., Okazaki, M., Matsuo, Y.: Earthquake shakes Twitter users: real-time
event detection by social sensors. In: Proceedings of WWW 2010, pp. 851–860
(2010)
17. Takamura, H., Inui, T., Okumura, M.: Extracting semantic orientations of words
using spin model. In: Proceedings of ACL 2005, pp. 133–140 (2005)
18. Varga, I., Sano, M., Torisawa, K., Hashimoto, C., Ohtake, K., Kawai, T., Oh, J.H.,
De Saeger, S.: Aid is out there: looking for help from tweets during a large scale
disaster. In: Proceedings of ACL 2013, pp. 1619–1629 (2013)
19. Vieweg, S., Hughes, A.L., Starbird, K., Palen, L.: Microblogging during two nat-
ural hazards events: what Twitter may contribute to situational awareness. In:
Proceedings of CHI 2010, pp. 1079–1088 (2010)
 A Deep Neural Architecture
for Sentence-Level Sentiment
Classification in Twitter Social
Networking

Huy Nguyen(B) and Minh-Le Nguyen

Japan Advanced Institute of Science and Technology, Ishikawa, Japan

{huy.nguyen,nguyenml}@jaist.ac.jp

Abstract. This paper introduces a novel deep learning framework

including a lexicon-based approach for sentence-level prediction of senti-
ment label distribution. We propose to first apply semantic rules and then
use a Deep Convolutional Neural Network (DeepCNN) for character-
level embeddings in order to increase information for word-level embed-
ding. After that, a Bidirectional Long Short-Term Memory network (Bi-
LSTM) produces a sentence-wide feature representation from the word-
level embedding. We evaluate our approach on three twitter sentiment
classification datasets. Experimental results show that our model can
improve the classification accuracy of sentence-level sentiment analysis
in Twitter social networking.

Keywords: Twitter · Sentiment classification · Deep learning

1 Introduction
Twitter sentiment classification have intensively researched in recent years
[6,14]. Different approaches were developed for Twitter sentiment classification
by using machine learning such as Support Vector Machine (SVM) with rule-
based features [15] and the combination of SVMs and Naive Bayes (NB) [17].
In addition, hybrid approaches combining lexicon-based and machine learning
methods also achieved high performance described in [13]. However, a problem
of traditional machine learning is how to define a feature extractor for a specific
domain in order to extract important features.
Deep learning models are different from traditional machine learning methods
in that a deep learning model does not depend on feature extractors because fea-
tures are extracted during training progress. The use of deep learning methods
becomes to achieve remarkable results for sentiment analysis [4,10,20]. Some
researchers used Convolutional Neural Network (CNN) for sentiment classifi-
cation. CNN models have been shown to be effective for NLP. For example,
[10] proposed various kinds of CNN to learn sentiment-bearing sentence vec-
tors, [4] adopted two CNNs in character-level to sentence-level representation
c Springer Nature Singapore Pte Ltd. 2018
K. Hasida and W. P. Pa (Eds.): PACLING 2017, CCIS 781, pp. 15–27, 2018.
https://doi.org/10.1007/978-981-10-8438-6_2
16 H. Nguyen and M.-L. Nguyen

for sentiment analysis. [20] constructs experiments on a character-level CNN for

several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is
another state-of-the-art semantic composition model for sentiment classification
with many variants described in [5]. The studies reveal that using a CNN is useful
in extracting information and finding feature detectors from texts. In addition,
a LSTM can be good in maintaining word order and the context of words. How-
ever, in some important aspects, the use of CNN or LSTM separately may not
capture enough information.
Inspired by the models above, the goal of this research is using a Deep Con-
volutional Neural Network (DeepCNN) to exploit the information of characters
of words in order to support word-level embedding. A Bi-LSTM produces a
sentence-wide feature representation based on these embeddings. The Bi-LSTM
is a version of [7] with Full Gradient described in [8]. In addition, the rules-
based approach also effects classification accuracy by focusing on important sub-
sentences expressing the main sentiment of a tweet while removing unnecessary
parts of a tweet. The paper makes the following contributions:

– We construct a tweet processor removing unnecessary sub-sentences from

tweets in order that the model learns important information in a tweet. We
share ideas with [1,6], however, our tweet processor keeps emoticons in tweets
and only uses rules to remove non-essential parts for handling negation.
– We train DeepCNN with Wide convolution on top of character embeddings
to produce feature maps which capture the morphological and shape infor-
mation of a word. The morphological and shape information illustrate how
words are formed, and their relationship to other words. DeepCNN trans-
forms the character-level embeddings into global fixed-sized feature vectors
at higher abstract level. Such character feature vectors contribute enriching
the information of words in a sentence.
– We create an integration of global fixed-size character feature vectors and
word-level embedding for the Bi-LSTM. The Bi-LSTM connects the informa-
tion of words in a sequence and maintains the order of words for sentence-level
representation.

The organization of the present paper is as follows: In Sect. 2, we describe

the model architecture which introduces the structure of the model. We explain
the basic idea of model and the way of constructing the model. Section 3 shows
results and analysis and Sect. 4 summarizes this paper.

2 Model Architecture

2.1 Basic Idea

Our proposed model consists in a deep learning classifier and a tweet processor.
The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The
tweet processor standardizes tweets and then applies rules called semantic rules
on datasets. We construct a framework that treats the deep learning classifier and
 A Deep Neural Architecture for Sentence-Level Sentiment Classification 17

Fig. 1. The overview of a deep learning system.

the tweet processor as two distinct components. We believe that standardizing

data is an important step to achieve high classification accuracy. To formulate
our problem in increasing the accuracy of the classifier, we illustrate our model
in Fig. 1 as follows:

1. Tweets are firstly considered via a processor based on preprocessing steps [6]
and the semantic rules-based method [1] in order to standardize tweets and
capture only important information containing the main sentiment of a tweet.
2. We use DeepCNN with Wide convolution for character-level embeddings. A
wide convolution can learn to recognize specific n-grams at every position in a
word that allows features to be extracted independently of these positions in
the word. These features maintain the order and relative positions of charac-
ters. A DeepCNN is constructed by two wide convolution layers and the need
of multiple wide convolution layers is widely accepted that a model construct-
ing by multiple processing layers have the ability to learn representations of
data with higher levels of abstraction [11]. Therefore, we use DeepCNN for
character-level embeddings to support morphological and shape information
for a word. The DeepCNN produces N global fixed-sized feature vectors for
N words.
3. A combination of the global fixed-size feature vectors and word-level embed-
ding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level represen-
tation by maintaining the order of words.

Our work is philosophically similar to [4]. However, our model is distinguished

with their approaches in two aspects:

– Using DeepCNN with two wide convolution layers to increase representation

with multiple levels of abstraction.
18 H. Nguyen and M.-L. Nguyen

– Integrating global character fixed-sized feature vectors with word-level

embedding to extract a sentence-wide feature set via Bi-LSTM. This deals
with three main problems: (i) Sentences have any different size; (ii) The
semantic and the syntactic of words in a sentence are captured in order to
increase information for a word; (iii) Important information of characters that
can appear at any position in a word are extracted.

In sub-section B, we introduce various kinds of dataset. The modules of our

model are constructed in other sub-sections.

2.2 Data Preparation

– Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains
1,600 K training tweets collected by a crawler from [6]. [6] constructed a test
set manually with 177 negative and 182 positive tweets. The Stanford test
set is small. However, it has been widely used in different evaluation tasks
[2,4,6].
– Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified
tweets collected by using search terms: #apple, #google, #microsoft and
#twitter. We construct the dataset as [3] for binary classification.
– Health Care Reform (HCR): This dataset was constructed by crawling
tweets containing the hashtag #hcr [16]. Task is to predict positive/negative
tweets [3].

2.3 Preprocessing

We firstly take unique properties of Twitter in order to reduce the feature space
such as Username, Usage of links, None, URLs and Repeated Letters. We then
process retweets, stop words, links, URLs, mentions, punctuation and accentu-
ation. For emoticons, [6] revealed that the training process makes the use of
emoticons as noisy labels and they stripped the emoticons out from their train-
ing dataset because [6] believed that if we consider the emoticons, there is a
negative impact on the accuracies of classifiers. In addition, removing emoticons
makes the classifiers learns from other features (e.g. unigrams and bi-grams)
presented in tweets and the classifiers only use these non-emoticon features to
predict the sentiment of tweets. However, there is a problem is that if the test
set contains emoticons, they do not influence the classifiers because emoticon
features do not contain in its training data. This is a limitation of [6], because
the emoticon features would be useful when using deep learning for classifying
the test data. Deep learning model captures emoticons as features and models
syntactic context information for words. Therefore, we keep emoticon features in
the datasets because deep learning models can capture more information from
emoticon features for increasing classification accuracy.
 A Deep Neural Architecture for Sentence-Level Sentiment Classification 19

Table 1. Semantic rules [1]

Rule Semantic rules Example - STS corpus Output

R11 If a sentence contains “but”, disregard @kirstiealley my she’s expen-
all previous sentiment and only take dentist is great but sive...=(
the sentiment of the part after “but” she’s expensive...=(
R12 If a sentence contains “despite”, only I’m not dead despite I’m not dead
take sentiment of the part before rumours to the
“despite” contrary
R13 If a sentence contains “unless”, and laptop charger is laptop
“unless” is followed by a negative broken - unless a little charger is
clause, disregard the “unless” clause cricket set up home broken
inside it overnight.
Typical at the worst
possible time
R14 If a sentence contains “while”, My throat is killing me, I still feel like
disregard the sentence following the and While I got a I’m about to
“while” and take the sentiment only decent night’s sleep last fall over
of the sentence that follows the one night, I still feel like
after the “while” I’m about to fall over
R15 If the sentence contains “however”, @lonedog it was quite
disregard the sentence preceding the bwahahah...you are the letdown
“however” and take the sentiment amazing! However, it
only of the sentence that follows the was quite the letdown
“however”

2.4 Semantic Rules (SR)

In Twitter social networking, people express their opinions containing sub-
sentences. These sub-sentences using specific PoS particles (Conjunction and
Conjunctive adverbs), like “but, while, however, despite, however” have differ-
ent polarities. However, the overall sentiment of tweets often focus on certain
sub-sentences. For example:
– @lonedog bwahahah...you are amazing! However, it was quite the letdown.
– @kirstiealley my dentist is great but she’s expensive...=(
In two tweets above, the overall sentiment is negative. However, the main
sentiment is only in the sub-sentences following but and however. This inspires
a processing step to remove unessential parts in a tweet. Rule-based approach
can assists these problems in handling negation and dealing with specific PoS
particles led to effectively affect the final output of classification [1,18]. [1] sum-
marized a full presentation of their semantic rules approach and devised ten
semantic rules in their hybrid approach based on the presentation of [18]. We
use five rules in the semantic rules set because other five rules are only used to
compute polarity of words after POS tagging or Parsing steps. We follow the
same naming convention for rules utilized by [1] to represent the rules utilized in
Another random document with
no related content on Scribd:
and over the brig o’ his nose; giving him a wauff, outlandish, and
rather blackguard sort of appearance, so that I was a thocht uneasy
at what neebours might surmeese concerning our intimacy; but the
honest man accounted for the thing in a very feasible manner, from
the falling down on that side of his head of one of the brass
candlesticks, while he was lying on his braidside, before ane of the
furms in the stramash.
His purpose of calling was to tell me that he couldna leave the
town without looking in upon me to bid me fareweel; mair betoken,
as he intended sending in his son Mungo wi’ the carrier for a trial, to
see how the line of life pleased him, and how I thocht he wad answer
—a thing which I was glad came from his side of the house, being
likely to be in the upshot the best for both parties. Yet I thocht he
wad find our way of doing so canny and comfortable, that it wasna
very likely he could ever start objections; and I must confess, that I
lookit forrit with nae sma’ degree of pride, seeing the probability of
my sune having the son of a Lammermuir farmer sitting cross-leggit,
cheek for jowl wi’ me, on the board, and bound to serve me at all
lawful times, by night and day, by a regular indenture of five years.
Maister Glen insisted on the laddie having a three months’ trial; and
then, after a wee show of standing out, just to make him aware that I
could be elsewhere fitted if I had a mind, I agreed that the request
was reasonable, and that I had nae yearthly objections to conforming
wi’t. So, after giein’ him his meridian, and a bit of shortbread, we
shook hands, and parted in the understanding, that his son would
arrive on the tap of limping Jamie the carrier’s cart, in the course,
say, of a fortnight.
Through the hale course of the forepart of the day, I remained
geyan queerish, as if something was working about my inwards, and
a droll pain atween my een. The wife saw the case I was in, and
advised me, for the sake of the fresh air, to take a step into the bit
garden, and try a hand at the spade, the smell of the fresh earth being
likely to operate as a cordial; but na—it wadna do; and whan I came
in at ane o’clock to my dinner, the steam of the fresh broth, instead
of making me feel as usual as hungry as a hawk, was like to turn my
stamach, while the sight of the sheep’s-head, ane o’ the primest anes
I had seen the hale season, made me as sick as a dog; so I could dae
naething but take a turn out again, and swig awa’ at the sma’ beer
that never seemed able to slocken my drouth. At lang and last, I
mindit having heard Andrew Redbeak, the excise-offisher, say, that
naething ever pat him right after a debosh, except something they ca’
a bottle of soda-water; so my wife dispatched Benjie to the place
where he kent it could be found, and he returned in a jiffie with a
thing like a blacking-bottle below his daidly, as he was bidden. There
being a wire ower the cork, for some purpose or ither, or maybe just
to look neat, we had some fight to get it torn away, but at last we
succeeded. I had turned about for a jug, and the wife was rummaging
for a screw, while Benjie was fiddling away wi’ his fingers at the cork
—sauf us! a’ at ance it gaed a thud like thunder, driving the cork ower
puir Benjie’s head, while it spouted up in his een like a fire-engine,
and I had only just time to throw down the jug, and up with the
bottle to my mouth. Luckily, for the sixpence it cost, there was a drap
o’t left, which tasted by all the world just like brisk dish-washings;
but, for a’ that, it had a wonderful power of setting me to rights; and
my noddle in a while began to clear up, like a March-day after a
heavy shower.
I mind very weel too, on the afternoon of the dividual day, that my
doorneebour, Thomas Burlings, pappit in; and, in our twa-handit
crack ower the counter, after asking me in a dry, curious way, if I had
come by nae skaith in the business of the play, he said, the thing had
now spread far and wide, and was making a great noise in the world.
I thocht the body a thocht sharp in his observes; so I pretended to
take it quite lightly, proceeding in my shaping-out a pair of buckskin-
breeches, which I was making for ane of the duke’s huntsmen; so,
seeing he was aff the scent, he said in a mair jocose way—“Weel,
speaking about buckskins, I’ll tell ye a gude story about that.”
“Let us hear’t,” said I; for I was in that sort of queerish way, that I
didna care muckle about being very busy.
“Ye’se get it as I heard it,” quo’ Thomas; “and it’s no less worth
telling, that it bears a gude moral application in its tail, after the
same fashion that a blister does gude by sucking away the vicious
humours of the body, thereby making the very pain it gies precious.”
And here—though maybe it was just my thocht—the body strokit his
chin, and gied me a kind of half glee, as muckle as saying, “take that
to ye, neebour.” But I deserved it all, and couldna take it ill aff his
hand, being, like mysel, ane of the elders of our kirk, and an honest
enough, preceese-speaking man.
“Ye see, ye ken,” said Thomas, “that the Breadalbane Fencibles, a
wheen Highland birkies, were put into camp on Fisherraw links,
maybe for the benefit of their douking, on account of the fiddle[14]—or
maybe in case the French should land at the water-mouth—or maybe
to gie the regiment the benefit of the sea air—or maybe to make their
bare houghs hardier, for it was the winter time, frost and snaw being
as plenty as ye like, and no sae scarce as pantaloons among the core,
or for some ither reason, gude, bad, or indifferent, which disna
muckle matter. But, ye see, the lang and the short o’ the story is, that
there they were encamped, man and mother’s son of them, going
through their dreels by day, and sleeping by night—the privates in
their tents, and the offishers in their markees; living in the course of
nature on their usual rations of beef and tammies, and sae on. So, ye
understand me, there was nae such smart orderings of things in the
army in thae days, the men not having the beef served out to them by
a butcher, supplying each company or companies by a written
contract, drawn up between him and the paymaster before sponsible
witnesses; but ilka ane bringing what pleased him, either tripe,
trotters, steaks, cow’s-cheek, pluck, hough, spar-rib, jiggot, or so
forth.”
14. See Dr Jamieson’s “Scottish Dictionary.”
“’Od!” said I, “Thomas, ye crack like a minister. Where did ye
happen to pick up all that knowledge?”
“Where should I have got it? but from an auld half-pay sergeant-
major, that lived in our spare room, and had been out in the
American war, having seen a power of service, and been twice
wounded,—ance in the aff cuit, and the ither time in the cuff of the
neck.”
“I thocht as muckle,” said I; “but say on, man; it’s unco
entertaining.”
“Weel,” continued he, “let me see where I was at when ye stoppit
me; for maybe I’ll hae to begin at the beginning again. For gif ye
yenterrupt me, or edge in a word, or put me out by asking questions,
I lose the thread of my discourse, and canna proceed.”
 “Ou, let me see,” said I, “ye was about the contract concerning the
beef.”
“Preceesely,” quo’ Thomas, stretching out his forefinger; “ye’ve
said it to a hair. At that time, as I was observing, the butcher didna
supply a company or companies, according to the terms of a
contract, drawn up before sponsible witnesses, between him and the
paymaster; but the soldiers got beef-money along with their pay;
with which said money, given them, ye observe, for said purpose,
they were bound and obligated, in terms of the statute, to buy,
purchase, and provide the said beef, twice a week or oftener, as it
might happen; an orderly offisher making inspection of the camp-
kettles regularly every forenoon at ane o’clock or thereabouts.
“So, as ye’ll pay attention to observe, there was a private in Captain
M‘Tavish’s company, the second to the left of the centre, of the name
of Duncan MacAlpine, a wee, hardy, blackavised, in-knee’d creature,
remarkable for naething that ever I heard tell of, except being
reported to have shotten a gauger in Badenoch, or thereabouts; and
for having a desperate red nose, the effects, ye observe, I daursay,—
the effects of drinking malt speerits.
“Weel, week after week passed ower, and better passed ower, and
Duncan played aff his tricks, like anither Herman Boaz, the slight-o’-
hand juggler—him that’s suspecket to be in league and paction with
the deil. But ye’ll hear.”
“’Od, it’s diverting, Thomas,” said I to him; “gang on, man.”
“Weel, ye see, as I was observing. Let me see, where was I at? Ou
ay, having a paction wi’ the deil. So, when all were watching beside
the camp-kettles, some stirring them wi’ spurtles, or parritch-sticks,
or forks, or whatever was necessary, the orderly offisher made a
point and practice of regularly coming by, about the chap of ane past
meridian, as I observed to ye before, to make inspection of what ilka
ane had wared his pay on; and what he had got simmering in the het
water for his dinner.
“So, on the day concerning which I am about to speak, it fell out, as
usual, that he happened to be making his rounds, halting a moment
—or twa, maybe—before ilka pat; the man that had the charge
thereof, by the way of stirring like, clapping down his lang fork, and
bringing up the piece of meat, or whatever he happened to be making
kail of, to let the inspector see whether it was lamb, pork, beef,
mutton, or veal. For, ye observe,” continued Thomas, gieing me, as I
took it to mysel, anither queer side look, “the purpose of the offisher
making the inspection, was to see that they laid out their pay-money
conform to military regulation; and no to filling their stamicks, and
ruining baith soul and body, by throwing it away on whisky, as but
ower mony, that aiblins should hae kent better, have dune but ower
aften.”
“’Tis but too true,” said I till him; “but the best will fa’ intil a faut
sometimes. We have a’ our failings, Thomas.”
“Just so,” answered Thomas; “but where was I at? Ou, about the
whisky. Weel, speaking about the whisky: ye see, the offisher,
Lovetenant Todrick, I b’lief they called him, had made an observe
about Duncan’s kettle; so, when he cam to him, Duncan was sitting
in the lown side of a dyke, with his red nose, and a pipe in his cheek,
on a big stane, glowering frae him anither way; and, as I was saying,
when he cam to him he said, ‘Weel, Duncan MacAlpine, what have ye
in your kettle the day, man?’
“And Duncan, rinning down his lang fork, answered in his ain
Highland brogue way—‘Please your honour, just my auld fav’rite,
tripe.’
“’Deed, Duncan,” said Lovetenant Todrick, or whatever they ca’d
him, “it is an auld fav’rite, surely, for I have never seen ye have
onything else for your denner, man.”
“Every man to his taste, please your honour,” answered Duncan
MacAlpine; “let ilka ane please her nainsel,”—hauling up a screed
half a yard lang; “ilka man to his taste, please your honour,
Lovetenant Todrick.”
“’Od, man,” said I to him; “’od, man, ye’re a deacon at telling a
story. Ye’re a queer hand. Weel, what cam next?”
“What think ye should come next?” quo’ Thomas, drily.
“I’m sure I dinna ken,” answered I.
“Weel,” said he, “I’ll tell; but where was I at?”
“Ou, at the observe of Lovetenant Todrick, or what they ca’ed him,
about the tripe; and the answer of Duncan MacAlpine on that head,
that ‘ilka man had his ain taste.’”
 “‘Vera true,’ said Lovetenant Todrick; ‘but lift it out a’thegither on
that dish, till I get my specs on; for never since I was born, did I ever
see before boiled tripe with buttons and button-holes intil’t.’”
At this I set up a loud laughing, which I couldna help, though it
was like to split my sides; but Thomas Burlings bade me whisht till I
heard him out.
“‘Buttons and button-holes!’ quo’ Duncan MacAlpine. ‘Look again,
wi’ yer specs; for ye’re surely wrang, Lovetenant Todrick.’”
“Buttons and button-holes! and ’deed I am surely right, Duncan,’
answered Lovetenant Todrick, taking his specs deliberately aff the
brig o’ his nose, and faulding them thegither, as he put them, first
into his morocco case, and syne into his pocket. ‘Howsomever,
Duncan MacAlpine, I’ll pass ye ower for this time, gif ye take my
warning, and for the future ware yer paymoney on wholesome
butcher’s meat, like a Christian, and no be trying to delude your ain
stamick, and your offisher’s een, by haddin’ up, on a fork, such a
heathenish make-up for a dish, as the leg of a pair o’ buckskin
breeches!’”
“Buckskin breeches!” said I; “and did he really and actually boil
siccan trash to his dinner?”
“Nae sae far south as that yet, friend,” answered Thomas. “Duncan
wasna sae bowed in the intellect as ye imagine, and had some spice
of cleverality about his queer manœuvres.—Eat siccan trash to his
dinner! Nae mair, Mansie, than ye intend to eat that iron guse ye’re
rinning alang that piece claith; but he wantit to make his offishers
believe that his pay gaed the right way—like the Pharisees of old that
keepit praying, in ell-lang faces, about the corners of the streets, and
gaed hame wi’ hearts full of wickedness and a’ manner of cheatrie.”
“And what way did his pay gang then?” askit I; “and hoo did he
live?”
“I telled ye before, frien,” answered Thomas, “that he was a
deboshed creature; and, like ower mony in the warld, likit weel what
didna do him ony good. It’s a wearyfu’ thing that whisky. I wish it
could be banished to Botany Bay.”
“It is that,” said I. “Muckle and nae little sin does it breed and
produce in this world.”
 “I’m glad,” quo’ Thomas, stroking down his chin in a slee way. “I’m
glad the guilty should see the folly o’ their ain ways: it’s the first step,
ye ken, till amendment;—and indeed I tell’t Maister Wiggie, when he
sent me here, that I could almost become gude for yer being mair
wary o’ yer conduct for the future time to come.”
This was like a thunder-clap to me, and I didna ken, for a jiffy,
what to feel, think, or do, mair than perceiving that it was a piece of
devilish cruelty on their pairts, taking things on this strict. As for
myself, I could freely take sacred oath on the Book, that I hadna had
a dram in my head for four months before; the knowledge of which
made my corruption rise like lightning, as a man is aye brave when
he is innocent; so, giein’ my pow a bit scart, I said briskly, “So ye’re
after some session business in this veesit, are ye?”
“Ye’ve just guessed it,” answered Thomas Burlings, sleeking down
his front hair with his fingers, in a sober way; “we had a meeting this
forenoon; and it was resolved ye should stand a public rebuke in the
meeting-house, on Sunday next.”
“Hang me, if I do!” answered I, thumping my nieve down with all
my might on the counter, and throwing back my cowl behind me,
into a corner.
“No, man!” added I, snapping with great pith my finger and thumb
in Thomas’s een; “no for all the ministers and elders that ever were
cleckit. They may do their best; and ye may tell them sae if ye like. I
was born a free man; I live in a free country; I am the subject of a free
king and constitution; and I’ll be shot before I submit to such rank
diabolical papistry.”
“Hooly and fairly,” quo’ Thomas, staring a wee astonished like, and
not a little surprised to see my birse up in this manner; for, when he
thought upon shearing a lamb, he fund he had catched a tartar; so,
calming down as fast as ye like, he said—“Hooly and fairly, Mansie”
(or Maister Wauch, I believe, he did me the honour to ca’ me),
“they’ll maybe no be sae hard as they threaten. But ye ken, my friend,
I’m speaking to ye as a brither; it was an unco-like business for an
elder, not only to gang till a play, which is ane of the deevil’s
rendezvouses, but to gang there in a state of liquor; making yoursel a
warld’s wonder—and you an elder of our kirk!—I put the question to
yourself soberly?”
 His threatening I could despise, and could have fought, cuffed, and
kickit, wi’ a’ the ministers and elders of the General Assembly, to say
naething of the Relief Synod, and the Burgher Union, before I wad
demeaned mysel to yield to what my inward speerit plainly telled me
to be rank cruelty and injustice; but ah! his calm, britherly, flattering
way I couldna thole wi’, and the tears came rapping into my een
faster than it cared my manhood to let be seen; so I said till him,
“Weel, weel, Thomas, I ken I have dune wrang; and I am sorry for’t—
they’ll never find me in siccan a scrape again.”
Thomas Burlings then cam forrit in a friendly way, and shook
hands wi’ me; telling that he wad go back and plead afore them in my
behalf. He said this ower again, as we pairted, at my shop door; and,
to do him justice, surely he hadna been waur than his word, for I
have aye attended the kirk as usual, standing, whan it came to my
rotation, at the plate, and naebody, gentle nor semple, ever spoke to
me on the subject of the playhouse, or minted the matter of the
rebuke from that day to this.
 ELPHIN IRVING, THE FAIRIES’
CUPBEARER.

By Allan Cunningham.

Chapter I.
The romantic vale of Corriewater, in Annandale, is regarded by the
inhabitants, a pastoral and unmingled people, as the last border
refuge of those beautiful and capricious beings, the fairies. Many old
people, yet living, imagine they have had intercourse of good words
and good deeds with the “gude folk;” and continue to tell that in the
ancient days the fairies danced on the hill, and revelled in the glen,
and showed themselves, like the mysterious children of the Deity of
old, among the sons and daughters of men. Their visits to the earth
were periods of joy and mirth to mankind, rather than of sorrow and
apprehension. They played on musical instruments of wonderful
sweetness and variety of note, spread unexpected feasts, the
supernatural flavour of which overpowered on many occasions the
religious scruples of the Presbyterian shepherds, performed
wonderful deeds of horsemanship, and marched in midnight
processions, when the sound of their elfin minstrelsy charmed
youths and maidens into love for their persons and pursuits; and
more than one family of Corriewater have the fame of augmenting
the numbers of the elfin chivalry. Faces of friends and relatives, long
since doomed to the battle trench, or the deep sea, have been
recognised by those who dared to gaze on the fairy march. The maid
has seen her lost lover, and the mother her stolen child; and the
courage to plan and achieve their deliverance has been possessed by,
at least, one border maiden. In the legends of the people of
Corrievale, there is a singular mixture of elfin and human adventure,
and the traditional story of the Cupbearer to the Queen of the Fairies
appeals alike to our domestic feelings and imagination.
In one of the little green loops or bends, on the banks of
Corriewater, mouldered walls, and a few stunted wild plum-trees and
vagrant roses, still point out the site of a cottage and garden. A well
of pure spring-water leaps out from an old tree-root before the door;
and here the shepherds, shading themselves in summer from the
influence of the sun, tell to their children the wild tale of Elphin
Irving and his sister Phemie; and, singular as the story seems, it has
gained full credence among the people where the scene is laid.
“I ken the tale and the place weel,” interrupted an old woman,
who, from the predominance of scarlet in her apparel, seemed to
have been a follower of the camp; “I ken them weel, and the tale’s as
true as a bullet to its aim, and a spark to powder. Oh, bonnie
Corriewater! a thousand times have I pu’ed gowans on its banks wi’
ane that lies stiff and stark on a foreign shore in a bloody grave:” and
sobbing audibly, she drew the remains of a military cloak over her
face, and allowed the story to proceed.
When Elphin Irving and his sister Phemie were in their sixteenth
year (for tradition says they were twins), their father was drowned in
Corriewater, attempting to save his sheep from a sudden swell, to
which all mountain streams are liable; and their mother, on the day
of her husband’s burial, laid down her head on the pillow, from
which, on the seventh day, it was lifted to be dressed for the same
grave. The inheritance left to the orphans may be briefly described:
seventeen acres of plough and pasture land, seven milk cows, and
seven pet sheep (many old people take delight in odd numbers); and
to this may be added seven bonnet pieces of Scottish gold, and a
broadsword and spear, which their ancestor had wielded with such
strength and courage in the battle of Dryfe-sands, that the minstrel
who sang of that deed of arms ranked him only second to the Scotts
and the Johnstones.
The youth and his sister grew in stature and in beauty. The brent
bright brow, the clear blue eye, and frank and blithe deportment of
the former, gave him some influence among the young women of the
valley; while the latter was no less the admiration of the young men,
and at fair and dance, and at bridal, happy was he who touched but
her hand, or received the benediction of her eye. Like all other
Scottish beauties, she was the theme of many a song; and while
tradition is yet busy with the singular history of her brother, song has
taken all the care that rustic minstrelsy can of the gentleness of her
spirit, and the charms of her person.
“Now I vow,” exclaimed a wandering piper, “by mine own
honoured instrument, and by all other instruments that ever yielded
music for the joy and delight of mankind, that there are more bonnie
songs made about fair Phemie Irving than about all the other
maidens of Annandale, and many of them are both high and bonnie.
A proud lass maun she be, if her spirit hears; and men say the dust
lies not insensible of beautiful verse; for her charms are breathed
through a thousand sweet lips, and no farther gone than yestermorn,
I heard a lass singing on a green hillside what I shall not readily
forget. If ye like to listen, ye shall judge; and it will not stay the story
long nor mar it much, for it is short, and about Phemie Irving.” And
accordingly he chanted the following rude verses, not
unaccompanied by his honoured instrument, as he called his pipe,
which chimed in with great effect, and gave richness to a voice which
felt better than it could express:—
FAIR PHEMIE IRVING.
I.

Gay is thy glen, Corrie,

With all thy groves flowering:
Green is thy glen, Corrie,
When July is showering;
And sweet is yon wood, where
The small birds are bowering,
And there dwells the sweet one
Whom I am adoring.

II.

Her round neck is whiter

Than winter when snowing;
Her meek voice is milder
Than Ae in its flowing;
The glad ground yields music
 Where she goes by the river;
One kind glance would charm me
For ever and ever.

III.

The proud and the wealthy

To Phemie are bowing;
No looks of love win they
With sighing or suing;
Far away maun I stand
With my rude wooing,
She’s a flow’ret too lovely
To bloom for my pu’ing—

IV.

O were I yon violet

On which she is walking;
O were I yon small bird
To which she is talking;
Or yon rose in her hand,
With its ripe ruddy blossom;
Or some pure gentle thought,
To be blest with her bosom!

This minstrel interruption, while it established Phemie Irving’s

claim to grace and to beauty, gave me additional confidence to
pursue the story.
But minstrel skill and true love tale seemed to want their usual
influence, when they sought to win her attention; she was only
observed to pay most respect to those youths who were most beloved
by her brother; and the same hour that brought these twins to the
world, seemed to have breathed through them a sweetness and an
affection of heart and mind, which nothing could divide. If, like the
virgin queen of the immortal poet, she walked “in maiden meditation
fancy free,” her brother Elphin seemed alike untouched with the
charms of the fairest virgins in Corrie. He ploughed his field, he
reaped his grain, he leaped, he ran and wrestled, and danced and
sang, with more skill and life and grace than all other youths of the
district; but he had no twilight and stolen interviews. When all other
young men had their loves by their side, he was single, though not
unsought; and his joy seemed never perfect save when his sister was
near him. If he loved to share his time with her, she loved to share
her time with him alone, or with the beasts of the field, or the birds
of the air. She watched her little flock late, and she tended it early;
not for the sordid love of the fleece, unless it was to make mantles for
her brother, but with the look of one who had joy in its company. The
very wild creatures, the deer and the hares, seldom sought to shun
her approach, and the bird forsook not its nest, nor stinted its song,
when she drew nigh; such is the confidence which maiden innocence
and beauty inspire.
It happened one summer, about three years after they became
orphans, that rain had been for a while withheld from the earth; the
hillsides began to parch, the grass in the vales to wither, and the
stream of Corrie was diminished between its banks to the size of an
ordinary rill. The shepherds drove their flocks to moorlands, and
marsh and tarn had their reeds invaded by the scythe, to supply the
cattle with food. The sheep of his sister were Elphin’s constant care;
he drove them to the moistest pastures during the day, and he often
watched them at midnight, when flocks, tempted by the sweet dewy
grass, are known to browse eagerly, that he might guard them from
the fox, and lead them to the choicest herbage. In these nocturnal
watchings he sometimes drove his little flock over the water of
Corrie, for the fords were hardly ankle-deep; or permitted his sheep
to cool themselves in the stream, and taste the grass which grew
along the brink. All this time not a drop of rain fell, nor did a cloud
appear in the sky.
One evening during her brother’s absence with the flock, Phemie
sat at her cottage door, listening to the bleatings of the distant folds,
and the lessened murmur of the water of Corrie, now scarcely
audible beyond its banks. Her eyes, weary with watching along the
accustomed line of road for the return of Elphin, were turned on the
pool beside her, in which the stars were glimmering fitful and faint.
As she looked, she imagined the water grew brighter and brighter; a
wild illumination presently shone upon the pool, and leaped from
bank to bank, and, suddenly changing into a human form, ascended
the margin, and passing her, glided swiftly into the cottage. The
visionary form was so like her brother in shape and air, that, starting
up, she flew into the house, with the hope of finding him in his
customary seat. She found him not; and impressed with the terror
which a wraith or apparition seldom fails to inspire, she uttered a
shriek so loud and so piercing as to be heard at Johnstonebank, on
the other side of the vale of Corrie.
An old woman now rose suddenly from her seat in the window-sill,
the living dread of shepherds, for she travelled the country with a
brilliant reputation for witchcraft, and thus she broke in upon the
narrative: “I vow, young man, ye tell us the truth upset and
downthrust; I heard my douce grandmother say that on the night
when Elphin Irving, disappeared—disappeared, I shall call it, for the
bairn can but be gone for a season, to return to us in his own
appointed time,—she was seated at the fireside at Johnstonebank;
the laird had laid aside his bonnet to take the Book, when a shriek
mair loud, believe me, than a mere woman’s shriek,—and they can
shriek loud enough, else they’re sair wranged,—came over the water
of Corrie, so sharp and shrilling, that the pewter plates dinnelled on
the wall; such a shriek, my douce grandmother said, as rang in her
ear till the hour of her death, and she lived till she was aughty and
aught, forty full ripe years after the event. But there is another
matter, which, doubtless, I cannot compel ye to believe; it was the
common rumour that Elphin Irving came not into the world like the
other sinful creatures of the earth, but was one of the Kane-bairns of
the fairies, whilk they had to pay to the enemy of man’s salvation
every seventh year. The poor lady-fairy,—a mother’s aye a mother, be
she elf’s flesh or Eve’s flesh,—hid her elf son beside the christened
flesh in Marion Irving’s cradle, and the auld enemy lost his prey for a
time. Now hasten on with your story, which is not a bodle the waur
for me. The maiden saw the shape of her brother, fell into a faint or a
trance, and the neighbours came flocking in. Gang on wi’ your tale,
young man, and dinna be affronted because an auld woman helped
ye wi’ it.”
It is hardly known, I resumed, how long Phemie Irving continued
in a state of insensibility. The morning was far advanced, when a
neighbouring maiden found her seated in an old chair, as white as
monumental marble; her hair, about which she had always been
solicitous, loosened from its curls, and hanging disordered over her
neck and bosom, her hands and forehead. The maiden touched the
one and kissed the other; they were as cold as snow; and her eyes,
wide open, were fixed on her brother’s empty chair, with the
intensity of gaze of one who had witnessed the appearance of a spirit.
She seemed insensible of any one’s presence, and sat fixed, and still,
and motionless. The maiden, alarmed at her looks, thus addressed
her: “Phemie, lass, Phemie Irving! Dear me, but this is awful! I have
come to tell ye that seven o’ yer pet sheep have escaped drowning in
the water; for Corrie, sae quiet and sae gentle yestreen, is rolling and
dashing frae bank to bank this morning. Dear me, woman, dinna let
the loss o’ the world’s gear bereave ye of your senses. I would rather
make ye a present of a dozen mug-ewes of the Tinwald brood mysel;
and now I think on’t, if ye’ll send ower Elphin, I will help him hame
with them in the gloaming mysel. So Phemie, woman, be comforted.”
At the mention of her brother’s name, she cried out, “Where is he?
oh, where is he?”—gazed wildly round, and, shuddering from head to
foot, fell senseless on the floor. Other inhabitants of the valley,
alarmed by the sudden swell of the river, which had augmented to a
torrent deep and impassable, now came in to inquire if any loss had
been sustained, for numbers of sheep and teds of hay had been
observed floating down about the dawn of the morning. They
assisted in reclaiming the unhappy maiden from her swoon; but
insensibility was joy compared to the sorrow to which she awakened.
“They have ta’en him away, they have ta’en him away;” she
chanted in a tone of delirious pathos; “him that was whiter and fairer
than the lily on Lyddal-lee. They have long sought, and they have
long sued, and they had the power to prevail against my prayers at
last. They have ta’en him away; the flower is plucked from among the
weeds, and the dove is slain amid a flock of ravens. They came with
shout, and they came with song, and they spread the charm, and they
placed the spell, and the baptised brow has been bowed down to the
unbaptised hand. They have ta’en him away, they have ta’en him
away; he was too lovely, and too good, and too noble, to bless us with
his continuance on earth; for what are the sons of men compared to
him?—the light of the moonbeam to the morning sun; the glow-
worm to the eastern star. They have ta’en him away, the invisible
dwellers of the earth. I saw them come on him, with shouting and
with singing, and they charmed him where he sat, and away they
bore him; and the horse he rode was never shod with iron, nor
owned before the mastery of human hand. They have ta’en him away,
over the water, and over the wood, and over the hill. I got but ae look
o’ his bonnie blue ee, but ae look. But as I have endured what never
maiden endured, so will I undertake what never maiden undertook,
—I will win him from them all. I know the invisible ones of the earth;
I have heard their wild and wondrous music in the wild woods, and
there shall a christened maiden seek him and achieve his
deliverance.”
She paused, and glancing round a circle of condoling faces, down
which the tears were dropping like rain, said, in a calm, but still
delirious tone,—
“Why do you weep, Mary Halliday? and why do you weep, John
Graeme? Ye think that Elphin Irving,—oh, it’s a bonnie, bonnie
name, and dear to many a maiden’s heart as well as mine,—ye think
that he is drowned in Corrie, and ye will seek in the deep, deep pools
for the bonnie, bonnie corse, that ye may weep over it, as it lies in its
last linen, and lay it, amid weeping and wailing, in the dowie
kirkyard. Ye may seek, but ye shall never find; so leave me to trim up
my hair, and prepare my dwelling, and make myself ready to watch
for the hour of his return to upper earth.”
And she resumed her household labours with an alacrity which
lessened not the sorrow of her friends.
 Chapter II.

Meanwhile, the rumour flew over the vale that Elphin Irving was
drowned in Corriewater. Matron and maid, old man and young,
collected suddenly along the banks of the river, which now began to
subside to its natural summer limits, and commenced their search;
interrupted every now and then by calling from side to side, and
from pool to pool, and by exclamations of sorrow for this misfortune.
The search was fruitless: five sheep, pertaining to the flock which he
conducted to pasture, were found drowned in one of the deep eddies;
but the river was still too brown, from the soil of its moorland
sources, to enable them to see what its deep shelves, its pools, and its
overhanging and hazelly banks concealed. They remitted further
search till the stream should become pure; and old man taking old
man aside, began to whisper about the mystery of the youth’s
disappearance: old women laid their lips to the ears of their coevals,
and talked of Elphin Irving’s fairy parentage, and his having been
dropped by an unearthly hand into a Christian cradle. The young
men and maids conversed on other themes; they grieved for the loss
of the friend and the lover, and while the former thought that a heart
so kind and true was not left in the vale, the latter thought, as
maidens will, on his handsome person, gentle manners, and merry
blue eye, and speculated with a sigh on the time when they might
have hoped a return for their love. They were soon joined by others
who had heard the wild and delirious language of his sister: the old
belief was added to the new assurance, and both again commented
upon by minds full of superstitious feeling, and hearts full of
supernatural fears, till the youths and maidens of Corrievale held no
more love trysts for seven days and nights, lest, like Elphin Irving,
they should be carried away to augment the ranks of the
unchristened chivalry.
It was curious to listen to the speculations of the peasantry. “For
my part,” said a youth, “if I were sure that poor Elphin escaped from
that perilous water, I would not give the fairies a pound of hiplock
wool for their chance of him. There has not been a fairy seen in the
land since Donald Cargill, the Cameronian, conjured them into the
Solway for playing on their pipes during one of his nocturnal
preachings on the hip of the Burnswark hill.”
“Preserve me, bairn,” said an old woman, justly exasperated at the
incredulity of her nephew, “if ye winna believe what I both heard and
saw at the moonlight end of Craigyburnwood on a summer night,
rank after rank of the fairy folk, ye’ll at least believe a douce man and
a ghostly professor, even the late minister of Tinwaldkirk; his only
son (I mind the lad weel, with his long yellow locks and his bonnie
blue eyes, when I was but a gilpie of a lassie), he was stolen away
from off the horse at his father’s elbow, as they crossed that false and
fearsome water, even Locherbriggflow, on the night of the
Midsummer Fair of Dumfries. Ay, ay, who can doubt the truth of
that? Have not the godly inhabitants of Almsfieldtown and
Timwaldkirk seen the sweet youth riding at midnight, in the midst of
the unhallowed troop, to the sound of flute and of dulcimer; and
though meikle they prayed, naebody tried to achieve his
deliverance?”
“I have heard it said, by douce folk and sponsible,” interrupted
another, “that every seven years the elves and fairies pay kane, or
make an offering of one of their children to the grand enemy of
salvation, and that they are permitted to purloin one of the children
of men to present to the fiend; a more acceptable offering, I’ll
warrant, than one of their own infernal brood, that are Satan’s sib
allies, and drink a drop of the deil’s blood every May morning. And
touching this lost lad, ye all ken his mother was a hawk of an
uncannie nest, a second cousin of Kate Kimmer, of Barfloshan, as
rank a witch as ever rode on ragwort. Ay, sirs, what’s bred in the
bone is ill to come out o’ the flesh.”
On these and similar topics, which a peasantry full of ancient
tradition and enthusiasm and superstition, readily associate with the
commonest occurrences of life, the people of Corrievale continued to
converse till the fall of evening; when each seeking their home,
renewed again the wondrous subject, and illustrated it with all that
popular belief and poetic imagination could so abundantly supply.
The night which followed this melancholy day was wild with wind
and rain; the river came down broader and deeper than before, and
the lightning, flashing by fits over the green woods of Corrie, showed
the ungovernable and perilous flood sweeping above its banks. It
happened that a farmer, returning from one of the border fairs,
encountered the full swing of the storm; but, mounted on an
excellent horse, and mantled from chin to heel in a good gray plaid,
beneath which he had the farther security of a thick great-coat, he sat
dry in his saddle, and proceeded in the anticipated joy of a subsided
tempest, and a glowing morning sun. As he entered the long grove,
or rather remains of the old Galwegian forest, which lines for some
space the banks of the Corriewater, the storm began to abate, the
wind sighed milder and milder among the trees; and here and there a
star, twinkling momentarily through the sudden rack of the clouds,
showed the river raging from bank to brae. As he shook the moisture
from his clothes, he was not without a wish that the day would dawn,
and that he might be preserved on a road which his imagination
beset with greater perils than the raging river; for his superstitious
feeling let loose upon his path elf and goblin, and the current
traditions of the district supplied very largely to his apprehension the
ready materials of fear.
Just as he emerged from the wood, where a fine sloping bank,
covered with short green sward, skirts the limit of the forest, his
horse made a full pause, snorted, trembled, and started from side to
side, stooped his head, erected his ears, and seemed to scrutinize
every tree and bush. The rider, too, it may be imagined, gazed round
and round, and peered warily into every suspicious-looking place.
His dread of a supernatural visitation was not much allayed, when he
observed a female shape seated on the ground at the root of a huge
old oak tree, which stood in the centre of one of those patches of
verdant sward, known by the name of “fairy rings,” and avoided by
all peasants who wish to prosper. A long thin gleam of eastern
daylight enabled him to examine accurately the being who, in this
wild place and unusual hour, gave additional terror to this haunted
spot. She was dressed in white from the neck to the knees; her arms,
long, and round, and white, were perfectly bare; her head,
uncovered, allowed her long hair to descend in ringlet succeeding
ringlet, till the half of her person was nearly concealed in the fleece.
Amidst the whole, her hands were constantly busy in shedding aside
the tresses which interposed between her steady and uninterrupted
gaze, down a line of old road which winded among the hills to an
ancient burial-ground.
 As the traveller continued to gaze, the figure suddenly rose, and
wringing the rain from her long locks, paced round and round the
tree, chanting in a wild and melancholy manner an equally wild and
delirious song:—
THE FAIRY OAK OF CORRIEWATER.
I.
The small bird’s head is under its wing,
The deer sleeps on the grass;
The moon comes out, and the stars shine down,
The dew gleams like the glass:
There is no sound in the world so wide,
Save the sound of the smitten brass,
With the merry cittern and the pipe
Of the fairies as they pass.—
But oh! the fire maun burn and burn,
And the hour is gone, and will never return.

II.

The green hill cleaves, and forth, with a bound,

Come elf and elfin steed;
The moon dives down in a golden cloud,
The stars grow dim with dread;
But a light is running along the earth,
So of heaven’s they have no need:
O’er moor and moss with a shout they pass,
And the word is, spur and speed.—
But the fire maun burn, and I maun quake,
And the hour is gone that will never come back.

III.

And when they come to Craigyburn wood,

The Queen of the Fairies spoke:—
“Come, bind your steeds to the rushes so green,
And dance by the haunted oak:
I found the acorn on Heshbon-hill,
In the nook of a palmer’s poke,
A thousand years since; here it grows!”
And they danced till the greenwood shook.—
But oh! the fire, the burning fire,
The longer it burns, it but blazes the higher.