Deep Learning Approaches to Text Production 1st

Edition Shashi Narayan download

https://textbookfull.com/product/deep-learning-approaches-to-
text-production-1st-edition-shashi-narayan/

Download more ebook from https://textbookfull.com

We believe these products will be a great fit for you. Click
the link to download now, or visit textbookfull.com
to discover even more!

Programming Machine Learning From Coding to Deep

Learning 1st Edition Paolo Perrotta

https://textbookfull.com/product/programming-machine-learning-
from-coding-to-deep-learning-1st-edition-paolo-perrotta/

Machine Learning in the Oil and Gas Industry :

Including Geosciences, Reservoir Engineering, and
Production Engineering with Python Yogendra Narayan
Pandey
https://textbookfull.com/product/machine-learning-in-the-oil-and-
gas-industry-including-geosciences-reservoir-engineering-and-
production-engineering-with-python-yogendra-narayan-pandey/

Learning Tensorflow A Guide to Building Deep Learning

Systems 1st Edition Tom Hope

https://textbookfull.com/product/learning-tensorflow-a-guide-to-
building-deep-learning-systems-1st-edition-tom-hope/

Deep Learning for the Life Sciences Applying Deep

Learning to Genomics Microscopy Drug Discovery and More
1st Edition Bharath Ramsundar

https://textbookfull.com/product/deep-learning-for-the-life-
sciences-applying-deep-learning-to-genomics-microscopy-drug-
discovery-and-more-1st-edition-bharath-ramsundar/
Introduction to Deep Learning Using R: A Step-by-Step
Guide to Learning and Implementing Deep Learning Models
Using R Taweh Beysolow Ii

https://textbookfull.com/product/introduction-to-deep-learning-
using-r-a-step-by-step-guide-to-learning-and-implementing-deep-
learning-models-using-r-taweh-beysolow-ii/

Programming PyTorch for Deep Learning Creating and

Deploying Deep Learning Applications 1st Edition Ian
Pointer

https://textbookfull.com/product/programming-pytorch-for-deep-
learning-creating-and-deploying-deep-learning-applications-1st-
edition-ian-pointer/

Deep Learning Pipeline: Building a Deep Learning Model

with TensorFlow 1st Edition Hisham El-Amir

https://textbookfull.com/product/deep-learning-pipeline-building-
a-deep-learning-model-with-tensorflow-1st-edition-hisham-el-amir/

From Deep Learning to Rational Machines 1st Edition

Cameron J. Buckner

https://textbookfull.com/product/from-deep-learning-to-rational-
machines-1st-edition-cameron-j-buckner/

Approaches to Enhance Industrial Production of Fungal

Cellulases Manish Srivastava

https://textbookfull.com/product/approaches-to-enhance-
industrial-production-of-fungal-cellulases-manish-srivastava/

Series ISSN 1947-4040

NARAYAN • GARDENT
Series Editor: Graeme Hirst, University of Toronto

Deep Learning Approaches to Text Production

Shashi Narayan, University of Edinburgh
Claire Gardent, CNRS/LORIA, Nancy

DEEP LEARNING APPROACHES TO TEXT PRODUCTION

Text production has many applications. It is used, for instance, to generate dialogue turns from dialogue
moves, verbalise the content of knowledge bases, or generate English sentences from rich linguistic
representations, such as dependency trees or abstract meaning representations. Text production is also at
work in text-to-text transformations such as sentence compression, sentence fusion, paraphrasing, sentence
(or text) simplification, and text summarisation. This book offers an overview of the fundamentals of neural
models for text production. In particular, we elaborate on three main aspects of neural approaches to text
production: how sequential decoders learn to generate adequate text, how encoders learn to produce better
input representations, and how neural generators account for task-specific objectives. Indeed, each text-
production task raises a slightly different challenge (e.g, how to take the dialogue context into account when
producing a dialogue turn, how to detect and merge relevant information when summarising a text, or how
to produce a well-formed text that correctly captures the information contained in some input data in the
case of data-to-text generation). We outline the constraints specific to some of these tasks and examine
how existing neural models account for them. More generally, this book considers text-to-text, meaning-
to-text, and data-to-text transformations. It aims to provide the audience with a basic knowledge of neural
approaches to text production and a roadmap to get them started with the related work. The book is mainly
targeted at researchers, graduate students, and industrials interested in text production from different forms
of inputs.

ABOUT SYNTHESIS
This volume is a printed version of a work that appears in the Synthesis
Digital Library of Engineering and Computer Science. Synthesis lectures
provide concise original presentations of important research and
development topics, published quickly in digital and print formats. For

MORGAN & CLAYPOOL

more information, visit our website: http://store.morganclaypool.com

store.morganclaypool.com
Deep Learning Approaches
to Text Production
 Synthesis Lectures on Human
Language Technologies
Editor
Grame Hirst, University of Toronto
Synthesis Lectures on Human Language Technologies is edited by Graeme Hirst of the University
of Toronto. The series consists of 50- to 150-page monographs on topics relating to natural
language processing, computational linguistics, information retrieval, and spoken language
understanding. Emphasis is on important new techniques, on new applications, and on topics that
combine two or more HLT subfields.

Deep Learning Approaches to Text Production

Shashi Narayan and Claire Gardent
2020

Linguistic Fundamentals for Natural Language Processing II: 100 Essentials from
Semantics and Pragmatics
Emily M. Bender and Alex Lascarides
2019

Cross-Lingual Word Embeddings

Anders Søgaard, Ivan Vulić, Sebastian Ruder, and Manaal Faruqui
2019

Bayesian Analysis in Natural Language Processing, Second Edition

Shay Cohen
2019

Argumentation Mining
Manfred Stede and Jodi Schneider
2018

Quality Estimation for Machine Translation

Lucia Specia, Carolina Scarton, and Gustavo Henrique Paetzold
2018

Natural Language Processing for Social Media, Second Edition

Atefeh Farzindar and Diana Inkpen
2017
iv
Automatic Text Simplification
Horacio Saggion
2017

Neural Network Methods for Natural Language Processing

Yoav Goldberg
2017

Syntax-based Statistical Machine Translation

Philip Williams, Rico Sennrich, Matt Post, and Philipp Koehn
2016

Domain-Sensitive Temporal Tagging

Jannik Strötgen and Michael Gertz
2016

Linked Lexical Knowledge Bases: Foundations and Applications

Iryna Gurevych, Judith Eckle-Kohler, and Michael Matuschek
2016

Bayesian Analysis in Natural Language Processing

Shay Cohen
2016

Metaphor: A Computational Perspective

Tony Veale, Ekaterina Shutova, and Beata Beigman Klebanov
2016

Grammatical Inference for Computational Linguistics

Jeffrey Heinz, Colin de la Higuera, and Menno van Zaanen
2015

Automatic Detection of Verbal Deception

Eileen Fitzpatrick, Joan Bachenko, and Tommaso Fornaciari
2015

Natural Language Processing for Social Media

Atefeh Farzindar and Diana Inkpen
2015

Semantic Similarity from Natural Language and Ontology Analysis

Sébastien Harispe, Sylvie Ranwez, Stefan Janaqi, and Jacky Montmain
2015
 v
Learning to Rank for Information Retrieval and Natural Language Processing, Second
Edition
Hang Li
2014

Ontology-Based Interpretation of Natural Language

Philipp Cimiano, Christina Unger, and John McCrae
2014

Automated Grammatical Error Detection for Language Learners, Second Edition

Claudia Leacock, Martin Chodorow, Michael Gamon, and Joel Tetreault
2014

Web Corpus Construction

Roland Schäfer and Felix Bildhauer
2013

Recognizing Textual Entailment: Models and Applications

Ido Dagan, Dan Roth, Mark Sammons, and Fabio Massimo Zanzotto
2013

Linguistic Fundamentals for Natural Language Processing: 100 Essentials from

Morphology and Syntax
Emily M. Bender
2013

Semi-Supervised Learning and Domain Adaptation in Natural Language Processing

Anders Søgaard
2013

Semantic Relations Between Nominals

Vivi Nastase, Preslav Nakov, Diarmuid Ó Séaghdha, and Stan Szpakowicz
2013

Computational Modeling of Narrative

Inderjeet Mani
2012

Natural Language Processing for Historical Texts

Michael Piotrowski
2012

Sentiment Analysis and Opinion Mining

Bing Liu
2012
vi
Discourse Processing
Manfred Stede
2011

Bitext Alignment
Jörg Tiedemann
2011

Linguistic Structure Prediction

Noah A. Smith
2011

Learning to Rank for Information Retrieval and Natural Language Processing

Hang Li
2011

Computational Modeling of Human Language Acquisition

Afra Alishahi
2010

Introduction to Arabic Natural Language Processing

Nizar Y. Habash
2010

Cross-Language Information Retrieval

Jian-Yun Nie
2010

Automated Grammatical Error Detection for Language Learners

Claudia Leacock, Martin Chodorow, Michael Gamon, and Joel Tetreault
2010

Data-Intensive Text Processing with MapReduce

Jimmy Lin and Chris Dyer
2010

Semantic Role Labeling

Martha Palmer, Daniel Gildea, and Nianwen Xue
2010

Spoken Dialogue Systems

Kristiina Jokinen and Michael McTear
2009

Introduction to Chinese Natural Language Processing

Kam-Fai Wong, Wenjie Li, Ruifeng Xu, and Zheng-sheng Zhang
2009
 vii
Introduction to Linguistic Annotation and Text Analytics
Graham Wilcock
2009

Dependency Parsing
Sandra Kübler, Ryan McDonald, and Joakim Nivre
2009

Statistical Language Models for Information Retrieval

ChengXiang Zhai
2008
Copyright © 2020 by Morgan & Claypool

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means—electronic, mechanical, photocopy, recording, or any other except for brief quotations
in printed reviews, without the prior permission of the publisher.

Deep Learning Approaches to Text Production

Shashi Narayan and Claire Gardent
www.morganclaypool.com

ISBN: 9781681737584 paperback

ISBN: 9781681737591 ebook
ISBN: 9781681738215 epub
ISBN: 9781681737607 hardcover

DOI 10.2200/S00979ED1V01Y201912HLT044

A Publication in the Morgan & Claypool Publishers series

SYNTHESIS LECTURES ON HUMAN LANGUAGE TECHNOLOGIES

Lecture #44
Series Editor: Grame Hirst, University of Toronto
Series ISSN
Print 1947-4040 Electronic 1947-4059
Deep Learning Approaches
to Text Production

Shashi Narayan
University of Edinburgh

Claire Gardent
CNRS/LORIA, Nancy

SYNTHESIS LECTURES ON HUMAN LANGUAGE TECHNOLOGIES #44

M
&C Morgan & cLaypool publishers
/stu
ABSTRACT
Text production has many applications. It is used, for instance, to generate dialogue turns from
dialogue moves, verbalise the content of knowledge bases, or generate English sentences from
rich linguistic representations, such as dependency trees or abstract meaning representations.
Text production is also at work in text-to-text transformations such as sentence compression,
sentence fusion, paraphrasing, sentence (or text) simplification, and text summarisation. This
book offers an overview of the fundamentals of neural models for text production. In particu-
lar, we elaborate on three main aspects of neural approaches to text production: how sequential
decoders learn to generate adequate text, how encoders learn to produce better input repre-
sentations, and how neural generators account for task-specific objectives. Indeed, each text-
production task raises a slightly different challenge (e.g, how to take the dialogue context into
account when producing a dialogue turn, how to detect and merge relevant information when
summarising a text, or how to produce a well-formed text that correctly captures the information
contained in some input data in the case of data-to-text generation). We outline the constraints
specific to some of these tasks and examine how existing neural models account for them. More
generally, this book considers text-to-text, meaning-to-text, and data-to-text transformations.
It aims to provide the audience with a basic knowledge of neural approaches to text production
and a roadmap to get them started with the related work. The book is mainly targeted at re-
searchers, graduate students, and industrials interested in text production from different forms
of inputs.

KEYWORDS
text production, text generation, deep learning, neural networks, meaning-to-text,
data-to-text, text-to-text, recurrent neural networks, sequence-to-sequence models,
attention, copy, coverage, AMR generation, RDF generation, verbalise, simplifica-
tion, compression, paraphrasing, dialogue generation, summarisation, content se-
lection, adequacy, input understanding, sentence representation, document repre-
sentation, communication goals, deep generators, reinforcement learning, evalua-
tion, grammatical, fluent, meaning-preserving, BLEU, ROUGE, relevant, coher-
ent
 xi

To my family.
– Shashi

A mes trois rayons de soleil, Jennifer, Gabrielle, et Caroline.

– Claire
 xiii

Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiii

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 What is Text Production? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Generating Text from Meaning Representations . . . . . . . . . . . . . . . . . 1
1.1.2 Generating Text from Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.3 Generating Text from Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 What’s Not Covered? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Our Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

PART I Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Pre-Neural Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1 Data-to-Text Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Meaning Representations-to-Text Generation . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 Grammar-Centric Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.2 Statistical MR-to-Text Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Text-to-Text Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.1 Sentence Simplification and Sentence Compression . . . . . . . . . . . . . . 18
2.3.2 Document Summarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Deep Learning Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.1 Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
xiv
3.1.2 Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.3 LSTMs and GRUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.4 Word Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 The Encoder-Decoder Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Learning Input Representations with Bidirectional RNNs . . . . . . . . . 33
3.2.2 Generating Text Using Recurrent Neural Networks . . . . . . . . . . . . . . 35
3.2.3 Training and Decoding with Sequential Generators . . . . . . . . . . . . . . 39
3.3 Differences with Pre-Neural Text-Production Approaches . . . . . . . . . . . . . . . 40
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

PART II Neural Improvements . . . . . . . . . . . . . . . . . . . . 43

4 Generating Better Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1 Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2 Copy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3 Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5 Building Better Input Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

5.1 Pitfalls of Modelling Input as a Sequence of Tokens . . . . . . . . . . . . . . . . . . . . 58
5.1.1 Modelling Long Text as a Sequence of Tokens . . . . . . . . . . . . . . . . . . 58
5.1.2 Modelling Graphs or Trees as a Sequence of Tokens . . . . . . . . . . . . . 58
5.1.3 Limitations of Sequential Representation Learning . . . . . . . . . . . . . . 62
5.2 Modelling Text Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.1 Modelling Documents with Hierarchical LSTMs . . . . . . . . . . . . . . . 63
5.2.2 Modelling Document with Ensemble Encoders . . . . . . . . . . . . . . . . . 65
5.2.3 Modelling Document With Convolutional Sentence Encoders . . . . . 68
5.3 Modelling Graph Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3.1 Graph-to-Sequence Model for AMR Generation . . . . . . . . . . . . . . . 71
5.3.2 Graph-Based Triple Encoder for RDF Generation . . . . . . . . . . . . . . 74
5.3.3 Graph Convolutional Networks as Graph Encoders . . . . . . . . . . . . . . 75
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

6 Modelling Task-Specific Communication Goals . . . . . . . . . . . . . . . . . . . . . . . . 81

6.1 Task-Specific Knowledge for Content Selection . . . . . . . . . . . . . . . . . . . . . . . . 81
 xv
6.1.1 Selective Encoding to Capture Salient Information . . . . . . . . . . . . . . 82
6.1.2 Bottom-Up Copy Attention for Content Selection . . . . . . . . . . . . . . . 84
6.1.3 Graph-Based Attention for Salient Sentence Detection . . . . . . . . . . . 86
6.1.4 Multi-Instance and Multi-Task Learning for Content Selection . . . . 88
6.2 Optimising Task-Specific Evaluation Metric with Reinforcement Learning . 91
6.2.1 The Pitfalls of Cross-Entropy Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.2.2 Text Production as a Reinforcement Learning Problem . . . . . . . . . . . 96
6.2.3 Reinforcement Learning Applications . . . . . . . . . . . . . . . . . . . . . . . . 101
6.3 User Modelling in Neural Conversational Model . . . . . . . . . . . . . . . . . . . . . . 106
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

PART III Data Sets and Conclusion . . . . . . . . . . . . . . . 111

7 Data Sets and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.1 Data Sets for Data-to-Text Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.1.1 Generating Biographies from Structured Data . . . . . . . . . . . . . . . . . 113
7.1.2 Generating Entity Descriptions from Sets of RDF Triples . . . . . . . 114
7.1.3 Generating Summaries of Sports Games from Box-Score Data . . . 115
7.2 Data Sets for Meaning Representations to Text Generation . . . . . . . . . . . . . 116
7.2.1 Generating from Abstract Meaning Representations . . . . . . . . . . . . 117
7.2.2 Generating Sentences from Dependency Trees . . . . . . . . . . . . . . . . . 118
7.2.3 Generating from Dialogue Moves . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.3 Data Sets for Text-to-Text Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.3.1 Summarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.3.2 Simplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.3.3 Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.3.4 Paraphrasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.1 Summarising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.2 Overview of Covered Neural Generators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.3 Two Key Issues with Neural NLG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
8.4 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.5 Recent Trends in Neural NLG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
xvi
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

Authors’ Biographies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

 xvii

List of Figures
1.1 Input contents and communicative goals for text production . . . . . . . . . . . . . . . 2
1.2 Shallow dependency tree from generation challenge surface realisation task . . . 3
1.3 Example input from the SemEval AMR-to-Text Generation Task . . . . . . . . . . 4
1.4 E2E dialogue move and text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Data-to-Text example input and output pair . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 A Robocup input and output pair example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2 Data to text: A pipeline architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Simplifying a sentence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 A Sentence/Compression pair . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5 Abstractive vs. extractive summarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.6 A document/summary pair from the CNN/DailyMail data set . . . . . . . . . . . . 22
2.7 Key modules in pre-neural approaches to text production . . . . . . . . . . . . . . . . 24

3.1 Deep learning for text generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.2 Feed-forward neural network or multi-layer perceptron . . . . . . . . . . . . . . . . . . 26
3.3 Convolutional neural network for sentence encoding . . . . . . . . . . . . . . . . . . . . 28
3.4 RNNs applied to a sentence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.5 Long-range dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.6 Sketches of LSTM and GRU cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.7 Two-dimensional representation of word embeddings . . . . . . . . . . . . . . . . . . . 32
3.8 RNN-based encoder-decoder architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.9 The German word “zeit” with its two translations . . . . . . . . . . . . . . . . . . . . . . 34
3.10 Bidirectional RNNs applied to a sentence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.11 RNN decoding steps (Continues.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.12 (Continued.) RNN decoding steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.13 RNN decoding: conditional generation at each step . . . . . . . . . . . . . . . . . . . . . 38
xviii
4.1 Example input/output with missing, added, or repeated information . . . . . . . 47
4.2 Focusing on the relevant source word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3 Sketch of the attention mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.4 Example delexicalisations from the E2E and WebNLG data sets . . . . . . . . . . 52
4.5 Interactions between slot values and sentence planning . . . . . . . . . . . . . . . . . . 52
4.6 Example of generated text containing repetitions . . . . . . . . . . . . . . . . . . . . . . . 53
4.7 The impact of coverage on repetition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.8 Evolution of the DA vector as generation progresses . . . . . . . . . . . . . . . . . . . . 56

5.1 Bidirectional RNN modelling document as a sequence of tokens . . . . . . . . . . 59

5.2 Linearising AMR for text production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3 Linearising RDF to prepare input-output pairs for text production . . . . . . . . 61
5.4 Linearising dialogue moves for response generation . . . . . . . . . . . . . . . . . . . . . 61
5.5 Linearising Wikipedia descriptions for text generation . . . . . . . . . . . . . . . . . . 62
5.6 Hierarchical document representation for abstractive document
summarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.7 Multi-agent document representation for abstractive document
summarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.8 Communication among multiple agents, each encoding a paragraph . . . . . . . 66
5.9 Extractive summarisation with a hierarchical encoder-decoder model . . . . . . . 69
5.10 Graph-state LSTMs for text production from AMR graphs . . . . . . . . . . . . . . 72
5.11 Graph-triple encoder for text production from RDF triple sets . . . . . . . . . . . . 76
5.12 Graph convolutional networks for encoding a sentence . . . . . . . . . . . . . . . . . . 77

6.1 An example of abstractive sentence summarisation . . . . . . . . . . . . . . . . . . . . . . 82

6.2 Selective encoding for abstractive sentence summarisation . . . . . . . . . . . . . . . 83
6.3 Heat map learned with the selective gate mechanism . . . . . . . . . . . . . . . . . . . . 84
6.4 Two-step process for content selection and summary generation . . . . . . . . . . . 85
6.5 Graph-based attention to select salient sentences for abstractive
summarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.6 Generating biography from Wikipedia infobox . . . . . . . . . . . . . . . . . . . . . . . . 88
6.7 Example of word-property alignments for the Wikipedia abstract and facts . . 89
 xix
6.8 The exposure bias in cross-entropy trained models . . . . . . . . . . . . . . . . . . . . . . 93
6.9 Text production as a reinforcement learning problem . . . . . . . . . . . . . . . . . . . . 97
6.10 The curriculum learning in application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.11 Deep reinforcement learning for sentence simplification . . . . . . . . . . . . . . . . 101
6.12 Extractive summarisation with reinforcement learning . . . . . . . . . . . . . . . . . 104
6.13 Inconsistent responses generated by a sequence-to-sequence model . . . . . . . 107
6.14 Single-speaker model for response generation . . . . . . . . . . . . . . . . . . . . . . . . 108
6.15 Examples of speaker consistency and inconsistency in the speaker model . . . 108
6.16 Responses to “Do you love me?” from the speaker-addressee model . . . . . . . 109

7.1 Infobox/text example from the WikiBio data set . . . . . . . . . . . . . . . . . . . . . . 114

7.2 Example data-document pair from the extended WikiBio data set . . . . . . . . 114
7.3 Wikidata/text example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.4 Example data-document pair from the RotoWire data set . . . . . . . . . . . . . . 116
7.5 Example input and output from the SemEval AMR-to-Text Generation
Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.6 Example shallow input from the SR’18 data set . . . . . . . . . . . . . . . . . . . . . . . 118
7.7 Example instance from the E2E data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.8 Example summary from the NewsRoom data set . . . . . . . . . . . . . . . . . . . . . . 121
7.9 An abridged example from the XSum data set . . . . . . . . . . . . . . . . . . . . . . . . 122
7.10 PWKP complex and simplified example pairs . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.11 Newsela example simplifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.12 GigaWord sentence compression or summarisation example . . . . . . . . . . . . . 125
7.13 Sentence compression example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.14 Example of abstractive compression from Toutanova et al. [2016] . . . . . . . . 127
7.15 Example of abstractive compression from Cohn and Lapata [2008] . . . . . . . 127
7.16 Example paraphrase pairs from ParaNMT-50 . . . . . . . . . . . . . . . . . . . . . . . . 128
7.17 Examples from the Twitter News URL Corpus . . . . . . . . . . . . . . . . . . . . . . . 129
7.18 Paraphrase examples from PIT-2015 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.19 Paraphrase examples from the MSR corpus . . . . . . . . . . . . . . . . . . . . . . . . . . 130
 xxi

List of Tables
6.1 An abridged CNN article and its story highlights (Continues.) . . . . . . . . . . . . 95
6.1 (Continued.) An abridged CNN article and its story highlights . . . . . . . . . . . . 96

7.1 Summary of publicly available large corpora for summarisation . . . . . . . . . . 120

7.2 Data statistics assessing extractiveness of summarisation data sets . . . . . . . . . 123
7.3 Summary of publicly available large sentential paraphrase corpora . . . . . . . . 128
 xxiii

Preface
Neural methods have triggered a paradigm shift in text production by supporting two key fea-
tures. First, recurrent neural networks allow for the learning of powerful language models which
can be conditioned on arbitrarily long input and are not limited by the Markov assumption. In
practice, this proved to allow for the generation of highly fluent, natural sounding text. Second,
the encoder-decoder architecture provides a natural and unifying framework for all generation
tasks independent of the input type (data, text, or meaning representation). As shown by the
dramatic increase in the number of conference and journal submissions on that topic, these two
features have led to a veritable explosion of the field.
In this book, we introduce the basics of early neural text-production models and contrast
them with pre-neural approaches. We begin by briefly reviewing the main characteristics of
pre-neural text-production models, emphasising the stark contrast with early neural approaches
which mostly modeled text-production tasks independent of the input type and of the com-
municative goal. We then introduce the encoder-decoder framework where, first, a continuous
representation is learned for the input and, second, an output text is incrementally generated
conditioned on the input representation and on the representation of the previously generated
words. We discuss the attention, copy, and coverage mechanisms that were introduced to im-
prove the quality of generated texts. We show how text-production can benefit from better input
representation when the input is a long document or a graph. Finally, we motivate the need for
neural models that are sensitive to the current communication goal. We describe different vari-
ants of neural models with task-specific objectives and architectures which directly optimise
task-specific communication goals. We discuss generation from text, data, and meaning repre-
sentations, bringing various text-production scenarios under one roof to study them all together.
Throughout the book we provide an extensive list of references to support further reading.
As we were writing this book, the field had already moved on to new architectures and
models (Transformer, pre-training, and fine-tuning have now become the dominant approach),
and we discussed these briefly in the conclusion. We hope that this book will provide a useful
introduction to the workings of neural text production and that it will help newcomers from
both academia and industry quickly get acquainted with that rapidly expanding field.
We would like to thank several people who provided data or images, and authorization to
use them in this book. In particular, we would like to thank Abigail See for the pointer-generator
model, Asli Celikyilmaz for the diagrams of deep communicating paragraph encoders, Bayu Dis-
tiawan Trisedya for graph-triple encoders, Bernd Bohnet for an example from the 2018 surface
realisation challenge, Diego Marcheggiani for graph convolutional network (GCN) diagrams, Ji-
wei Tan for hierarchical document encoders and graph-based attention mechanism using them,
xxiv PREFACE
Jonathan May for an abstract meaning representation (AMR) graph, Laura Perez-Beltrachini
for an extended RotoWire example, Linfeng Song for graph-state long short-term memories
(LSTMs) for text production from AMR graphs, Marc’Aurelio Ranzato for exposure bias and
curriculum learning algorithm diagrams, Qingyu Zhou for selective encoding figures, Sam Wise-
man for a corrected RotoWire example, Sebastian Gehrmann for the bottom-up summariza-
tion diagram, Tsung-Hsien Wen for an alternative coverage mechanism plot, Xingxing Zhang
for reinforcement learning for sentence simplification, and Yannis Konstas for AMR-to-text and
data-to-text examples. Huge thanks to Emiel Krahmer, Graeme Hirst, and our anonymous re-
viewer for reviewing our book and providing us with detailed and constructive feedback. We
have attempted to address all the issues they raised. All the remaining typos and inadequacies
are entirely our responsibility. Finally, we would like to thank Morgan & Claypool Publishers
for working with us in producing this manuscript. A very special thanks goes to Michael Morgan
and Christine Kiilerich for always encouraging us and keeping us on track.

Shashi Narayan and Claire Gardent

March 2020
 1

CHAPTER 1

Introduction
In this chapter, we outline the differences between text production and text analysis, we introduce
the main text-production tasks this book is concerned with (i.e., text production from data, from
text, and from meaning representations) and we summarise the content of each chapter. We also
indicate what is not covered and introduce some notational conventions.

1.1 WHAT IS TEXT PRODUCTION?

While natural language understanding [NLU, Bates, 1995] aims to analyse text, text production,
or natural language generation [NLG, Gatt and Krahmer, 2018, Reiter and Dale, 2000], focuses
on generating texts. More specifically, NLG differs from NLU in two main ways (cf. Figure 1.1).
First, unlike text analysis, which always takes text as input, text production has many possible
input types, namely, text [e.g., Nenkova and McKeown, 2011], data [e.g., Wiseman et al., 2017],
or meaning representations [e.g., Konstas et al., 2017]. Second, text production has various
potential goals. For instance, the goal may be to summarise, verbalise, or simplify the input.
Correspondingly, text production has many applications depending on what the input
(data, text, or meaning representations) and what the goal is (simplifying, verbalising, para-
phrasing, etc.). When the input is text (text-to-text or T2T generation), text production can
be used to summarise the input document [e.g., Nenkova and McKeown, 2011], simplify a
sentence [e.g., Shardlow, 2014, Siddharthan, 2014] or respond to a dialogue turn [e.g., Mott
et al., 2004]. When the input is data, NLG can further be used to verbalise the content of a
knowledge [e.g., Power, 2009] or a database [e.g., Angeli et al., 2010], generate reports from
numerical [e.g., Reiter et al., 2005] or KB data [e.g., Bontcheva and Wilks, 2004], or generate
captions from images [e.g., Bernardi et al., 2016]. Finally, NLG has also been used to regenerate
text from the meaning representations designed by linguists to represent the meaning of natural
language [e.g., Song et al., 2017].
In what follows, we examine generation from meaning representations, data, and text in
more detail.

1.1.1 GENERATING TEXT FROM MEANING REPRESENTATIONS

There are two main motivations for generating text from meaning representations.
First, an algorithm that converts meaning representations into well-formed text is a nec-
essary component of traditional pipeline NLG systems [Gatt and Krahmer, 2018, Reiter and
Dale, 2000]. As we shall see in Chapter 2, such systems include several modules, one of them
2 1. INTRODUCTION
Verbalize
Respond
Summarize
Simplify
…

Communication Goal

Meaning Representations
- AMRs, Dialogue Moves, Abstract Dependency Trees …
Data
Input - Knowledge- and Data-Bases, Numerical, Graphical Data …
Text
- Sentence, Text, Multiple Texts, Dialogue Turn …

Figure 1.1: Input contents and communicative goals for text production.

(known as the surface realisation module) being responsible for generating text from some ab-
stract linguistic representation derived by the system. To improve reusability, surface realisation
challenges have recently been organised in an effort to identify input meaning representations
that could serve as a common standard for NLG systems, thereby fueling research on that topic.
Second, meaning representations can be viewed as an interface between NLU and NLG.
Consider translation, for instance. Instead of learning machine translation models, which di-
rectly translate surface strings into surface strings, an interesting scientific challenge would be to
learn a model that does something more akin to what humans seem to do, i.e., first, understand
the source text, and second, generate the target text from the conceptual representation issued
from that understanding (indeed a recent paper by Konstas et al. [2017] mentions this as future
work). A similar two-step process (first, deriving a meaning representation from the input text,
and second, generating a text from that meaning representation) also seems natural for such
tasks as text simplification or summarisation.
Although there are still relatively few approaches adopting a two-step interpret-and-
generate process or reusing existing surface realisation algorithms, there is already a large trend
of research in text production which focuses on generating text from meaning representations
produced by a semantic parser [May and Priyadarshi, 2017, Mille et al., 2018] or a dialogue
manager [Novikova et al., 2017b]. In the case of semantic parsing, the meaning representations
capture the semantics of the input text and can be exploited as mentioned above to model a
 1.1. WHAT IS TEXT PRODUCTION? 3

may

SBJ VC P

report be .

NMOD NMOD PRD

the troublesome deficit

AMOD NMOD NMOD NMOD APPO

most the august trade out

NMOD AMOD TMP

merchandise due tomorrow

Figure 1.2: Input shallow dependency tree from the generation challenge surface realisation
task for the sentence “The most troublesome report may be the August merchandise trade deficit due
out tomorrow.”

two-step process in applications such as simplification [Narayan and Gardent, 2014], summari-
sation [Liu et al., 2015] or translation [Song et al., 2019b]. In the case of dialogue, the input
meaning representation (a dialogue move) is output by the dialogue manager in response to the
user input and provides the input to the dialogue generator, the module in charge of generating
the system response.
While a wide range of meaning representations and syntactic structures have been pro-
posed for natural language (e.g., first-order logic, description logic, hybrid logic, derivation
rather than derived syntactic trees), three main types of meaning representations have recently
gained traction as input to text generation: meaning representations derived from syntactic de-
pendency trees (cf. Figure 1.2), meaning representations derived through semantic parsing (cf.
Figure 1.3), and meaning representations used as input to the generation of a dialogue engine
response (cf. Figure 1.4). All three inputs gave rise to shared tasks and international challenges.
The Surface Realisation shared task [Belz et al., 2012, Mille et al., 2018] focuses on gen-
erating sentences from linguistic representations derived from syntactic dependency trees and
includes a deep and a shallow track. For the shallow track, the input is an unordered, lemma-
tised syntactic dependency tree and the main focus is on linearisation (deriving the correct word
order from the input tree) and morphological inflection (deriving the inflection from a lemma
and a set of morphosyntactic features). For the deep track, on the other hand, the input is a de-
Another Random Scribd Document
with Unrelated Content
 The troops are said to have formed a huge semi-circle with the
G.P.O. as the centre, and, starting from the river, are driving the
rebels back street by street, till eventually they will be in a small
enclosure, when they will bombard it to pieces.
The G.P.O. has such valuable records, etc., and the contents of the
safes are so precious, that they will not raze it to the ground if they
can help it; but it has so much subterranean space, that would
afford cover to thousands of Sinn Feiners, that we hear they are
going to fire some "gas" shells into it and then rush it!
Up to yesterday afternoon they had got to Abbey Street on the
right, and no doubt were closing in equally on other sides. The shells
had started several fires; nearly all the shops on the quay on the
side of the Custom House were burning yesterday afternoon, and
later in the evening many others broke out.
I cannot give you any idea of what it was like when I went to bed.
I sent for Mrs. B., the manager's wife, such a splendid little woman,
and together we watched it from my window, which is high up and
looked in the right direction.
 It was the most awe-inspiring sight I have ever seen. It seemed
as if the whole city was on fire, the glow extending right across the
heavens, and the red glare hundreds of feet high, while above the
roar of the fires the whole air seemed vibrating with the noise of the
great guns and machine-guns. It was an inferno! We remained spell-
bound, and I can't tell you how I longed for you to see it. We had
only just come down from the window—we had been standing on
the window ledge leaning out—when H. came and told us no one
was to look out of the windows as there was cross-firing from the
United Service Club and another building, and Mr. O'B., who was
watching the fires from his window, had a bullet a few inches from
his head!!
About 2 a.m. I woke to find the room illuminated in spite of dark
blinds and curtains, and I rushed to the window and saw an
enormous fire; it seemed to be in the direction of the Four Courts,
which is in the hands of the Sinn Feiners, and we hear this morning
that a portion of the buildings was burnt last night.[A]
[A] This was incorrect; it was the Linen Hall barracks that were
burnt.

Yesterday Lord S. had a narrow escape from a sniper who has

been worrying this street for two days and could not be located. He
was picking off soldiers during the fighting in Grafton Street, but
later turned his attention to the cross streets between this and
Grafton Street, and there as nearly as possible got Lord S., who was
coming back to us from the Castle.
The military thought the man was on our roof, which made us all
bristle with indignation—the mere idea of the wretch being on our
hotel; but a thorough search proved he was not here, though he
evidently had access to some roof.
In this respect we are much better off than our friends the V.'s.
They came into their town house only about a month ago, and being
in Upper Mount Street it was in one of the most active haunts of the
snipers. They had several on their roof, and when they went up to
bed at night they could hear the snipers walking about and talking
on the roof. Does it not make one creep to think of it? Mr. V. had his
bed put on the upper landing exactly under the trap-door on to the
roof, so that had the rebels attempted to enter the house at night
they would have come down "plop" on to him in his bed. He
surrounded himself with all the arms he could muster, and the
wretched Mrs. V. lay in bed and quaked, expecting any minute to
hear a battle royal raging outside her bedroom door. In this street an
old lady of seventy-three was shot through the leg in her own room,
and was taken to Dr. W.'s home, where she had to have her leg
amputated; and in another house a servant flashed on her electric
light when going to bed and was instantly shot through the head!
Our friend Miss K. also had a narrow escape. She had only just left
her drawing-room, when a bullet passed straight through the room
and buried itself in a picture.
Yesterday afternoon, when the firing in Grafton Street was over,
the mob appeared and looted the shops, clearing the great provision
shops and others. From the back of this hotel you look down on an
alley that connects with Grafton Street,—and at the corner, the shop
front in Grafton Street, but with a side entrance into this lane, is a
very large and high-class fruiterer. From the windows we watched
the proceedings, and I never saw anything so brazen! The mob were
chiefly women and children, with a sprinkling of men. They swarmed
in and out of the side door bearing huge consignments of bananas,
the great bunches on the stalk, to which the children attached a cord
and ran away dragging it along. Other boys had big orange boxes
which they filled with tinned and bottled fruits. Women with their
skirts held up received showers of apples and oranges and all kinds
of fruit which were thrown from the upper windows by their pals;
and ankle-deep on the ground lay all the pink and white and silver
paper and paper shavings used for packing choice fruits. It was an
amazing sight, and nothing daunted these people. Higher up at
another shop we were told a woman was hanging out of a window
dropping down loot to a friend, when she was shot through the head
by a sniper, probably our man; the body dropped into the street and
the mob cleared. In a few minutes a hand-cart appeared and
gathered up the body, and instantly all the mob swarmed back to
continue the joyful proceedings!
H. and Lord S. were sitting at the window for a few minutes
yesterday when the fruit shop was being looted, and saw one of the
funniest sights they had ever seen. A very fat, very blousy old
woman emerged from the side street and staggered on to the
pavement laden with far more loot than she could carry. In her arms
she had an orange box full of fruit, and under her shawl she had a
great bundle tied up which kept slipping down. Having reached the
pavement, she put down her box and sat on it, and from her bundle
rolled forth many tins of fruit. These she surveyed ruefully, calling on
the Almighty and all the saints to help her!! From these she solemnly
made her selection, which she bound up in her bundle and hoisted,
with many groans and lamentations, on her back and made off with,
casting back many longing looks at the pile of things left on the
pavement, which were speedily disposed of by small boys.
On Wednesday when the looting was going on in Sackville Street a
fine, large boot shop was receiving attention from swarms of looters.
Ragged women and children were seen calmly sitting in the window
trying on boots and shoes, and one old woman with an eye to future
needs made up a bundle of assorted sizes and tied them up in her
apron. She had only reached the pavement, when she bethought her
to leave her bundle in a corner and return for a further consignment
which she tied up in her shawl. On returning to the street great was
her rage and indignation on finding the original bundle had
disappeared. Then were there sore lamentations and violent abuse
of the police, who could not even "protect the property of a poor old
woman."
In Sackville Street was a very large shop called Clery's; for some
reason the looters were afraid to start on it, and old women passed
up and down gazing longingly at fur coats and silken raiment and
saying sorrowfully, "Isn't Clery's broke yet?" and "Isn't it a great
shame that Clery's is not broke!" Humour and tragedy are so
intermixed in this catastrophe. A very delicate elderly lady who is
staying here said to me this morning, in answer to my inquiry as to
how she had slept: "I could not sleep at all. When the guns ceased
the awful silence made me so nervous!" I know exactly what she
meant. When the roar of the guns ceases you can feel the silence.
4 p.m.
When I had got so far this morning I got an urgent message from
the Red Cross asking me to make more armlets for the workers.
With two other ladies I had been making them yesterday, so I
collected my helpers and we worked till lunch, when another request
came that we would make four large Red Cross flags, as they were
going to try to bury some of the dead and needed the flags for the
protection of the parties. We have just finished them, and are
wondering what will be the next call. It is such a good thing I have
my sewing-machine here.
On Wednesday evening Lord S. was at Mercer's Hospital with a
doctor when eleven dead were brought in, and a priest brought in a
rifle he had taken from a dead Sinn Feiner. It had an inscription in
German and the name of the factory in Berlin, which Lord S. copied.
It is believed that nearly all the arms and ammunition are of German
make, and it is said that the cruiser that was sunk on Saturday was
bringing heavy guns and forty officers, but I don't know if there is
any truth in that. The opinion is very strong that the Sinn Feiners
were led to believe that they would have great German
reinforcements, and that all they had to do was to hold the troops
here for a couple of days while the Germans landed a big force on
the west coast of Ireland. We also hear that Sir R. Casement has
been shot in London, but you probably know a great deal more
about that than I do, as we see no papers and are completely cut off
from all news.
On Wednesday three of the ringleaders were caught, and it is said
they were shot immediately! It is also believed that Larkin was shot
on the top of a house in St. Stephen's Green, but as the rebels still
hold the house it has not been possible to identify him, but he is said
to have been here on Monday.[B]
[B] This was incorrect; it appears Larkin was not in Dublin.

5 p.m.
Colonel C. has just come in, having been in the thick of it for forty-
eight hours. He tells us the Post Office has been set on fire by the
Sinn Feiners, who have left it. If this is true, and it probably is, I fear
we have lost all our valuable possessions, including my diamond
pendant, which was in my jewel-case in H.'s safe.
To-day about lunch-time a horrid machine-gun suddenly gave
voice very near us. We thought it was in this street, but it may have
been in Kildare Street; also the sniper reappeared on the roofs, and
this afternoon was opposite my bedroom window judging from the
sound. I pulled down my blinds. A man might hide for weeks on the
roofs of these houses among the chimney stacks and never be found
as long as he had access to some house for food. When we were
working in my room this afternoon he fired some shots that could
not have been more than twenty yards away.
The serious problem of food is looming rather near, as nothing has
come into the city since Saturday. Boland's bakery, an enormous
building, is in the hands of the rebels, who have barricaded all the
windows with sacks of flour, and it is said it will have to be blown up.
There is not a chance of getting them out in any other way. The
rebels also have Jacob's biscuit factory, where there are still huge
stores of flour. Every prominent building and every strategic position
was taken before the authorities at the Castle woke to the fact that
there was a rebellion!
I was almost forgetting to tell you how splendidly one of H.'s men
behaved when the G.P.O. was taken. When the rebels took
possession they demanded the keys from the man who had them in
charge. He quietly handed over the keys, having first abstracted the
keys of H.'s room!
 Imagine such self-possession at such a terrible moment.
A young man has come to stay in the hotel who saw the taking of
the G.P.O. He was staying at the hotel exactly opposite the building
and went into the G.P.O. to get some stamps. As he was leaving the
office a detachment of about fifteen Irish Volunteers marched up
and formed up in front of the great entrance. He looked at them
with some curiosity, supposing they were going to hold a parade;
two more detachments arrived, and immediately the word of
command was given, and they rushed in through the door. Shots
were fired inside the building, and, as the young man said, he
"hooked it" back to the hotel, which was one of those burnt a few
days later. The whole thing occupied only a few moments, as, being
Bank Holiday, there was only a small staff in the building.
6.30 p.m. A party of soldiers and a young officer have just
arrived to search the roof for the sniper. They say he is on the roof
of the annexe, which is connected with the main building by
covered-in bridges. They are now on the roof and shots are being
fired, so I expect they have spotted him.
When N. was out last night another ambulance had a bad
experience. They had fetched three wounded Sinn Feiners out of a
house, and were taking them to hospital, when they came under
heavy fire. The driver was killed, so the man beside him took the
wheel and was promptly wounded in both legs. The car then ran
away and wrecked itself on a lamp post. Another ambulance had to
run the gauntlet and go to the rescue! On the whole as far as
possible the rebels have respected the Red Cross, but not the white
flag. In house-to-house fighting there can be no connected action,
and yesterday when a house was being stormed the rebels hung out
a white flag, and when the troops advanced to take them prisoners
they were shot down from a house a few doors higher up the street,
so now no more white flag signals are to be recognised. If they want
to surrender they must come out and take their own risks.
 We asked N. if he knew what had happened to the ambulance
that had two men missing yesterday, and he told us they were in the
act of entering a Sinn Fein house to bring out wounded with two
other men when the ambulance came under such heavy fire that, as
it contained one or two other wounded men, it had to beat a retreat
and moved off. Two of the volunteer helpers ran after it and
succeeded in reaching it and climbed in, but the other two took
refuge in the area, and N. did not know how or when they were
rescued. This is an instance of the extreme danger that attends the
ambulance work. The marvel is that the casualties are so few.
Guinness's Brewery have made three splendid armoured cars by
putting great long boilers six feet in diameter on to their large motor
lorries. Holes are bored down the sides to let in air, and they are
painted grey. The driver sits inside too. They each carry twenty-two
men or a ton of food in absolute security. N. saw them at the Castle
being packed with men; nineteen got in packed like herrings, and
three remained outside. Up came the sergeant: "Now then,
gentlemen, move up, move up: the car held twenty-two yesterday; it
must hold twenty-two to-day"; and in the unfortunate three were
stuffed. It must have been suffocating, but they were taken to their
positions in absolute safety.
Saturday, 29th, 10 a.m.
Last night was an agitating one. The sniper was very active, and
after dinner several shots struck the annexe, one or two coming
through the windows, and one broke the glass roof of the bridge. Mr.
B., who never loses his head, decided to get all the people out of the
annexe, with staff (about forty people); and all we in the main
building, whose rooms look out on the back, were forbidden to have
lights in our rooms at all. There was such a strong feeling of
uneasiness throughout the hotel, and always the danger of its being
set on fire, that about 10 p.m. H. said we must be prepared at any
moment to leave the hotel if necessary. So we went up to our room
and in pitch darkness groped about and collected a few things (F.'s
miniature and the presentation portrait of him, my despatch case
with his letters, my fur coat, hat and boots), and we took them
down to the sitting-room, which H. uses as an office, on the first
floor. All the people in the hotel were collected in the lounge, which
is very large and faces the street, and the whole of the back was in
complete darkness. The firing quieted down, and about 11.30 we
crept up to our room and lay down in our clothes. When dawn broke
I got up and undressed and had two hours' sleep. All the rest of the
guests spent the night in the lounge.
This morning we hear an officer has been to say that the shots
fired into the hotel last night were fired by the military. People were
constantly pulling up their blinds for a moment with the lights on to
look at the city on fire, and the military have orders to fire on
anything that resembles signalling without asking questions.
Reliable news has come in this morning that nothing remains of
the G.P.O. but the four main walls and the great portico. It is
absolutely burnt out. The fires last night were terrible, but we dared
not look out. Eason's Library and all the shops and buildings
between O'Connell Bridge and the G.P.O. on both sides of Sackville
Street are gone.
It is difficult to think of the position without intense bitterness,
though God knows it is the last thing one wishes for at such a time.
In pandering to Sir E. Carson's fanaticism and allowing him to raise a
body of 100,000 armed men for the sole purpose of rebellion and
provisional government the Government tied their own hands and
rendered it extremely difficult to stop the arming of another body of
men, known to be disloyal, but whose avowed reason was the
internal defence of Ireland! In Ulster the wind was sown, and, my
God, we have reaped the whirlwind!
We hear that many of our wounded are being sent to Belfast, as
the hospitals here are crowded, and the food problem must soon
become acute. Mr. O'B. told me his ambulance picked up four
wounded, three men and a woman, and took them to the nearest
hospital. The woman was dying, so they stopped at a church and
picked up a priest; arrived at the hospital the authorities said they
could not possibly take them in as they had not enough food for
those they had already taken, but when they saw the condition of
the woman they took her in to die, and the others had to be taken
elsewhere.
If the main walls of the G.P.O. remain standing it may be we shall
find the safe in H.'s room still intact. It was built into the wall, and
my jewel-case was in it, but all our silver, old engravings, and other
valuables were stored in the great mahogany cupboards when we
gave up our house in the autumn, as being the safest place in
Dublin.
4 p.m.
Sir M. N. has just rung up to say the rebels have surrendered
unconditionally. We have no details, and the firing continues in
various parts of the town. But if the leaders have surrendered it can
only be a question of a few hours before peace is restored, and we
can go forth and look on the wreck and desolation of this great city.
So ends, we hope, this appalling chapter in the history of Ireland
—days of horror and slaughter comparable only to the Indian Mutiny.
This seems a suitable place, dear G., to end this letter, and I hope to
start a happier one to-morrow.
Yours,
L. N.
 Third Letter
Sunday, April 30th, 10 a.m.
Dearest G.,—When I closed my letter last night with the news that
the rebel leaders had surrendered I hoped to start this new letter in
a more cheerful strain; but while we were dining last night H. was
rung up from the Castle to hear that the whole of Sackville Street
north of the G.P.O. right up to the Rotunda was on fire and blazing
so furiously that the fire brigade were powerless; nothing could go
near such an inferno. There was nothing to be done but let the fire
exhaust itself.
If this was true, it involved the loss of the Post Office Accountant's
Office opposite the G.P.O., the Sackville Street Club, Gresham and
Imperial Hotels, and other important buildings, and would have
increased H.'s difficulties enormously, as it would have been
necessary to build up the Post Office organisation again, with no
records, registers, accounts, or documents of any kind—at best a
stupendous task. However, fortunately this morning we hear the
reports were exaggerated. The Imperial Hotel, Clery's great shop,
and one or two others were burnt, but the upper part of the street
escaped, and the Accountant's Office and the Sackville Street Club
were not touched.
This morning Mr. C, who has been H.'s great support all through
this trying time (his second in command being away ill), and several
other members of the staff are coming here, and with H. they are
going down to see what remains of the G.P.O. It is being guarded
from looters, as, from the enormous number of telegraph
instruments destroyed, there must be a large quantity of copper and
other metal,—a very valuable asset,—and also several thousand
pounds in cash for payment of staff and soldiers' dependants,
besides heaps of other valuable property.
Here I must tell you how absolutely heroic the telephone staff
have been at the Exchange. It is in a building a considerable
distance from the G.P.O., and the Sinn Feiners have made great
efforts to capture it. The girls have been surrounded by firing; shots
have several times come into the switch-room, where the men took
down the boards from the back of the switch-boards and arranged
them as shelters over the girls' heads to protect them from bullets
and broken glass. Eight snipers have been shot on buildings
commanding the Exchange, and one of the guard was killed
yesterday; and these twenty girls have never failed. They have been
on duty since Tuesday, sleeping when possible in a cellar and with
indifferent food, and have cheerfully and devotedly stuck to their
post, doing the work of forty. Only those on duty on the outbreak of
the rebellion could remain; those in their homes could never get
back, so with the aid of the men who take the night duty these girls
have kept the whole service going. All telegrams have had to be sent
by 'phone as far as the railway termini, and they have simply saved
the situation. It has been magnificent!
The shooting is by no means over, as many of the Sinn Fein
strongholds refuse to surrender. Jacob's biscuit factory is very
strongly held, and when the rebels were called on to surrender they
refused unless they were allowed to march out carrying their arms!

As the book passes through the press, I learn on the

one unimpeachable authority that the story about
Messrs. Jacob & Co., however picturesque, is purely
apocryphal.
M.L.N.
The Sinn Fein Rebellion as I saw it, page 59.

It is said that when Jacob was told that the military might have to
blow up the factory he replied: "They may blow it to blazes for all I
care; I shall never make another biscuit in Ireland." I don't know if
this is true, but it very well may be, for he has been one of the
model employers in Dublin, and almost gave up the factory at the
time of the Larkin strike, and only continued it for the sake of his
people; and so it will be with the few great industries in the city.
Dublin is ruined.
Yesterday I made a joyful discovery. When we came back from
Italy in March, H. brought back from the office my large despatch-
case in which I keep all F.'s letters. I did not remember what else
was in it, so I investigated and found my necklet with jewelled cross
and the pink topaz set (both of these being in large cases would not
go in the jewel-case), also the large old paste buckle; so I am not
absolutely destitute of jewellery. But, best of all, there were the
three little handkerchiefs F. sent me from Armentières with my initial
worked on them; for these I was grieving more than for anything,
and when I found them the relief was so great I sat with them in my
hand and cried.
This week has been a wonderful week for N. Never before has a
boy of just seventeen had such an experience. Yesterday morning he
was at the Automobile Club filling cans of petrol from casks for the
Red Cross ambulance. He came in to lunch reeking of petrol. In the
afternoon he went round with the Lord Mayor in an ambulance
collecting food for forty starving refugees from the burnt-out district
housed in the Mansion House, and after tea went out for wounded
and brought in an old man of seventy-eight shot through the body.
He was quite cheery over it, and asked N. if he thought he would
recover. "Good Lord! yes; why not?" said N., and bucked the old man
up!
Some of the staff who came here this morning had seen a copy of
the Daily Mail yesterday, which devoted about six lines to the
condition of things in Ireland and spoke of a Sinn Fein riot in which
four soldiers and about six rebels had been killed. If that is all the
English people are being told of a rebellion which 30,000 troops and
many batteries of artillery are engaged in putting down, my letter
will be rather a surprise to you; and as the news must come out, the
English people will hardly be pleased at being kept in the dark. Such
a rebellion cannot be suppressed like a Zeppelin raid. During the first
three days our casualties were nearly 1,000; now we hear they are
close on 2,000.[C]
[C] This was exaggerated, our total casualties being about
1,380.

The College of Surgeons in St. Stephen's Green is still held by the

rebels, so the firing of machine-guns from the Shelbourne Hotel and
the United Service Club goes on as before, and there is intermittent
firing in all directions. I doubt if it will quite cease for some days, as
these strongholds will not surrender. Also the incendiary fires will
probably continue. The great fire in Sackville Street last night was no
doubt the work of incendiaries, as all the fires had died down. There
was no wind, no shells were being fired, and no reason for the
outbreak, but with all the relations and sympathisers of the rebels at
large the fires may very well continue.
The staff have just returned. They are quite unnerved by what
they have seen; they report nothing left of the G.P.O. but the four
outside walls and portico, so we have lost everything. They say it is
like a burned city in France.
May 1st, 11 a.m.
I had no time to continue this yesterday, but during the afternoon
three of the rebel strongholds surrendered—Jacob's, Boland's, and
the College of Surgeons on St. Stephen's Green. From this last
building 160 men surrendered and were marched down Grafton
Street. It is said that among them was Countess Markievicz, dressed
in a man's uniform. It is also said that the military made her take
down the green republican flag flying over the building herself and
replace it by a white one: when she surrendered she took off her
bandolier and kissed it and her revolver before handing them to the
officer. She has been one of the most dangerous of the leaders, and
I hope will be treated with the same severity as the men. People
who saw them marched down Grafton Street said they held
themselves erect, and looked absolutely defiant!
2 p.m.
To-day for the first time since Easter Monday the Irish Times
issued a paper with news of the rebellion. Very pluckily they had
brought out a paper on Tuesday, but it contained only the
proclamation and no reference to the rebellion, but a long account of
Gilbert and Sullivan's operas which were to have been performed
this week.
To-day's paper bears the dates "Friday, Saturday, and Monday,
April 28th, 29th, and May 1st"—an incident unique, I should think, in
the history of the paper.
It contains the various proclamations in full, which I will cut out
and send to you. Please keep them, as they will be of interest in the
future.
The paper states that Sir R. Casement is a prisoner in the Tower.
So he was not shot without trial, as we were told. It also gives a list
of the large shops and business establishments that have been
destroyed—a total of 146.
It really seemed delightful to hear the little paper boys calling their
papers about the streets again, and they had a ready sale for their
papers at three times their value. This so encouraged them that in
the afternoon they were running about again calling "Stop press."
Several people went out and bought papers, only to find they were
the same papers they had paid 3d. for in the morning.
"But this is the same paper I bought this morning."
"Sure, and it is, ma'am, but there's been a power of these papers
printed, and they're not going to print any more till they're all sold."
Another lady thought she would drive a lesson home, so she said:
"But you said it was a 'Stop press,' and you knew it was not."
 "It is, miss, but sure they hadn't time to print the 'stop press' on
it!!"
("Stop press" is the latest news, usually printed on the back of the
paper.)
Anyway, so great was the relief at seeing a paper again that no
one grudged the urchins their little harvest.
Yesterday H. visited the Telephone Exchange, and a point was
cleared up that has mystified everyone; and that is why, when the
rebels on Easter Monday took every building of importance and
every strategic position, did they overlook the Telephone Exchange?
Had they taken it we should have been absolutely powerless, unable
to send messages or telegrams for troops. The exchange is situated
in Crown Alley, off Dame Street, and the superintendent told H. an
extraordinary story. It seems when the rebels had taken the G.P.O.
they marched a detachment to take the exchange, when just as they
were turning into Crown Alley an old woman rushed towards them
with arms held up calling out, "Go back, boys, go back; the place is
crammed with military"; and supposing it to be in the hands of our
troops they turned back. This was at noon. At 5 p.m. our troops
arrived and took it over.
This saved the whole situation. Whether the woman was on our
side or whether she thought she had seen soldiers will never be
known.
When at the Castle yesterday H. got a copy of The Times for
Saturday, the first paper we have seen since Monday, so you can
imagine how eagerly we scanned the news about Ireland. More has
got out than we expected, but still nothing like the true position. We
rubbed our eyes when we read that "two battalions" had been sent
to Ireland, and wondered if it could possibly have been a printer's
error for two divisions (40,000 men) which actually arrived on
Wednesday. The people were in the streets of Kingstown for twenty
hours watching the troops pass through. Since then many more
troops and artillery have come in.
2 p.m.
I have just returned from walking round the G.P.O. and Sackville
Street with H. and some of the officials. It passes all my powers of
description, only one word describes it, "Desolation." If you look at
pictures of Yprès or Louvain after the bombardment it will give you
some idea of the scene.
We looked up through the windows of the G.P.O. and saw the safe
that was in H.'s room still in the wall, and the door does not appear
to have been opened or the safe touched, but the whole place has
been such an inferno one would think the door must have been red-
hot. Among all the débris the fire was still smouldering, and we
could not penetrate inside. I picked up a great lump of molten
metal, a fantastic shape with bits of glass embedded in it. It is bright
like silver, but they tell me it is lead. It is quite curious. Do you
realise, G., that out of all H.'s library he now does not possess a
single book, except one volume of his Dante, and I not even a silver
teaspoon!!
Everything belonging to F. has gone; as he gave his life in the war,
so an act of war has robbed us of everything belonging to him—our
most precious possession.
It has almost broken H. up; but he has no time to think, which is
perhaps a good thing.
The old Morland and Smith mezzotints have also gone—things we
can never replace.
Behind the G.P.O. was the Coliseum Theatre, now only a shell; and
on the other side of the street was the office of the Freeman's
Journal, with all the printing machinery lying among the débris, all
twisted and distorted; but, worst of all, behind that was a great
riding school, where all the horses were burnt to death.
 If at all possible you ought to come over for Whitsuntide. You will
see such a sight as you will never see in your life unless you go to
Belgium.
When we came here H. was scandalised at the condition of the
G.P.O. The whole frontage was given up to sorting offices, and the
public office was in a side street, a miserable, dirty little place, that
would have been a disgrace to a small country town.
H. found that plans had been drawn up and passed for the
complete reconstruction of the interior, building in a portion of the
courtyard an office for sorting purposes, leaving the frontage for the
public office with entrance under the great portico.
So H. hustled, and the work was completed and opened to the
public six weeks ago.
It was really beautiful. The roof was a large glass dome, with
elaborate plaster work, beautiful white pillars, mosaic floor, counters
all of red teak wood, and bright brass fittings everywhere—a public
building of which any great city might be proud; and in six weeks all
that is left is a smoking heap of ashes!
N. had an extraordinary find inside one of the rooms. About six
yards from the main wall he found, covered with ashes and a beam
lying across it, a motor cycle. It was lying on its side. He got it out
and found it perfect, tyres uninjured and petrol in the tank, and he
rode it to the hotel, and has now taken it to the Castle to hand over
to the police.
May 2nd, 10 a.m.
Last evening after tea I walked all round the ruined district with N.
and two ladies from the hotel. The streets were thronged with
people, and threading their way among the crowd were all sorts of
vehicles: carts carrying the bodies of dead horses that had been shot
the first day and lain in the streets ever since; fire brigade
ambulances, followed by Irish cars bringing priests and driven by fire
brigade men. Then motors with Red Cross emblems carrying white-
jacketed doctors would dart along, followed by a trail of Red Cross
nurses on bicycles, in their print dresses and white overalls, their
white cap-ends floating behind them, all speeding on their errand of
mercy to the stricken city.
From time to time we came across on the unwashed pavement
the large dark stain telling its own grim story, and in one place the
blood had flowed along the pavement for some yards and down into
the gutter; but enough of horrors. We came sadly back, and on the
steps we met Mr. O'B. returning from a similar walk. He could hardly
speak of it, and said he stood in Sackville Street and cried, and many
other men did the same.
Last night after dinner we were sitting in the room H. uses as a
temporary office overlooking the street, when firing began just
outside. They were evidently firing at the offices of the Sinn Fein
Volunteers at the bottom of the road. It was probably the last stand
of the rebels, and the firing was very sharp and quick. We thought
bullets must come into the hotel. I was reading aloud some bits out
of the Daily Mail, and the men were smoking. They moved my chair
back to the wall between the windows out of the line of fire; but the
firing became so violent we decided it was foolhardy to remain, so
we deserted the room, took our papers, and went and sat on the
stairs till it was over.
Since then we have not heard a shot fired; and it would seem that
as we were present at the first shots fired in Sackville Street on
Easter Monday so we have been present at the last fired eight days
later in Dawson Street.
Out of all the novel experiences of the last eight days two things
strike me very forcibly. The first is that, under circumstances that
might well have tried the nerves of the strongest, there has been no
trace of fear or panic among the people in the hotel, either among
the guests or staff. Anxiety for absent friends of whom no tidings
could be heard, though living only in the next square, one both felt
and heard; but of fear for their own personal safety I have seen not
one trace, and the noise of battle after the first two days seemed to
produce nothing but boredom. The other is a total absence of
thankfulness at our own escape.
It may come; I don't know. Others may feel it; I don't. I don't
pretend to understand it; but so it is. Life as it has been lived for the
last two years in the midst of death seems to have blunted one's
desire for it, and completely changed one's feelings towards the
Hereafter.
Now, G., I will end this long letter, and my next will probably deal
with normal if less interesting matters, but intense interest must
remain in the reconstruction of this great city.
Surely it must be possible to find men who will rule with firmness
and understanding this fine people—so kindly, so emotional, so
clever, so easily guided, and so magnificent when wisely led. One
prays they may be found, and found quickly, and that we may live to
see a Dublin restored to its former stateliness with a Government
worthy of the nation.
Ever yours,
L. N.
 Fourth Letter
Thursday, May 4th.
Dearest G.,—I had not intended writing again so soon, but things
are still happening that I think you will like to know, so I am going
on with this series of letters, though I don't know when you will get
them. But as by this time you will have seen N. you will have heard
many details from him. How much he will have to tell his school-
fellows when he returns to Shrewsbury to-morrow! I hoped to have
sent my second and third letters by N., and in fact had actually
packed them with his things. But when I told H. he said the rules
were so stringent about letters that N. would certainly be questioned
as to whether he was carrying any, and if he replied in the
affirmative, which he certainly would have done, the letters would
undoubtedly be confiscated and N. might get into serious trouble. So
I had to unpack them again and must keep them till the censorship
is removed, which will probably be in a few days. They have been
written under much stress of circumstances, and are the only record
we have of this most deeply interesting time, so I don't want to lose
them altogether.
I am not too well, as they say here. The loss of eight nights' sleep
seems to have robbed me of the power of sleeping for more than an
hour or two at a stretch, and even that is attended often with horrid
dreams and nightmares. But this is only the effect of over-strain, and
no doubt will pass, though my head feels like a feather bed; so don't
expect too much from these later letters.
Last night after dinner, when H. and I were sitting upstairs in
attendance on the telephone, who should walk in but Dr. W. We had
not met throughout the rebellion, so he had heaps to tell us. His
wife and children were down at Greystones, and the poor thing had
had a terribly anxious time, hearing nothing reliable of her husband
or of her father, Lord S. What she did hear was that Dr. W. had been
killed and also that H. had been shot in the G.P.O. She became so
anxious that her faithful Scotch nurse was determined to get into
Dublin and get news or die in the attempt. I must tell you her
adventures, not only to show you how impossible it was to get into
the city, but also it is such an extraordinary story of endurance and
devotion that it ought to be recorded.
The girl started from Greystones at 2.30 p.m. on the Thursday, I
think it was, carrying for the officers' home 14 lbs. of beef and 4 lbs.
of butter, as Mrs. W. feared supplies would have run short, since
nothing could be got in Dublin except at exorbitant prices (7s. a
dozen for eggs and 14s. for a pair of chickens); so the girl started
carrying a dead weight of 18 lbs.
She walked to Bray (five miles) and took train to Kingstown; here
she had to take to the road, as the line beyond Kingstown was
wrecked. She walked to Merrion Gates along the tram line about four
miles, when she was stopped by sentries. She retraced her steps as
far as Merrion Avenue (one mile), went up Merrion Avenue, and tried
the Stillorgan–Donnybrook route. Here she got as far as Leeson
Street Bridge (six miles), when she was within 300 yards of her
destination, Dr. W.'s house. Here again she was stopped by sentries
and turned back. She walked back to Blackrock (seven miles), when
she was again stopped by sentries. She then returned up Merrion
Avenue and, seeing that all routes were impossible to Dublin, took
the road to Killiney (five miles), where she arrived about 11.30 p.m.,
having done thirty miles. Here she got hospitality at a cottage and
stayed the remainder of the night there, paying for her
accommodation with the 4 lbs. of butter, but she stuck gamely to the
beef.
Next day she walked five miles to Shankhill, when she met a cart
going to Bray viâ Killiney, so she rode back to Killiney on it and from
thence to Bray. She then walked the five miles from Bray back to
Greystones, her starting point.
 Arrived back, she reached home absolutely exhausted, having
walked forty miles, and dropped down saying, "There's your beef,
and I never got there or heard anything." Mrs. W. was greatly
distressed at her having carried the meat back when so exhausted
and asked her why she had not given it away. "And what for should I
give it away when we'll be wanting it ourselves maybe?"
Next day Dr. W. managed to get a telephone message through to
his wife and relieved her anxiety.
He told us that on the first or second night of the rebellion—he
could not remember which—two ladies of the Vigilance Committee
patrolling the streets at night came on a soldier lying wounded in an
alley off Dawson Street, where he had crawled on being wounded.
They went to Mercer's Hospital and gave information, and stretcher-
bearers were sent out to bring in the man, the ladies accompanying
them. When he was on the stretcher the two ladies walked up to the
railings of St. Stephen's Green and gave the Sinn Feiners inside a
regular dressing down, telling them they were skunks and cowards
to shoot people down from behind bushes and asking them why
they did not come out and fight in the open like men. Meanwhile the
stretcher-bearers had taken the man to the hospital, where Dr. W.
saw him.
"Well, my man; where are you hurt?"
"Divil a pellet, sorr, above the knee," laughing.
"Does it pain you?"
"Not at all, sorr. Wait till I show you." He pulled up his trousers
and showed five bullet shots below the knee.
"What regiment?"
"Royal Irish, sorr, like Michael Cassidy, of Irish nationality; and I
bear no ill-will to nobody."
 Cheery soul! His great pride was that about forty shots had been
fired at him and not one hit him above the knee.
Dr. W. must bear a charmed life. He told us of several escapes he
had. One, the most dramatic, I must tell you.
You know he is one of the surgeons to Mercer's Hospital, and had
to be perpetually operating there at all hours of the day and night,
besides having his own private hospital, in which he takes wounded
officers. It too was filled with rebellion victims, so his work was
tremendous.
One night he left Mercer's about 1 a.m., accompanied by another
doctor. When passing in front of the Shelbourne Hotel they were
challenged by our troops there. On explaining who they were they
were of course allowed to proceed, and they stepped briskly out,
wanting to get home. Suddenly, on the same pavement, about
twenty yards away as far as they could judge in the black darkness,
out flashed two little lights from small electric lamps, evidently Sinn
Fein signals. Dr. W. stopped and said to his companion: "Did you see
that? it was a signal," when almost before the words were out of his
mouth two rifles blazed straight at them, almost blinding them with
the flash, and they felt the bullets whiz past their heads. The two
Sinn Feiners, having signalled, waited long enough to see if their
signal was returned, and then fired straight at where by their
footsteps they supposed Dr. W. and his friend to be, and missed
them by an inch or two.
Dr. W. and his friend got into the shelter of a doorway and
flattened themselves out, trying to look as if they were not there,
and quite forgetting that they both had lighted cigarettes, whose red
tips should have been a beacon light to a vital spot had the Sinn
Feiners noticed them. But for some reason they did not proceed
further, and Dr. W. heard their steps dying away in the distance.
Meanwhile his companion had his finger on the electric bell of the
doorway where they were hiding, and after a time which seemed like
an eternity an upper window opened and a voice inquired who was
there, whereupon the woman of the house came down and let them
in, and they spent the remainder of the night there.
Yesterday the Post Office was able to pay the separation
allowances to the soldiers' wives. Last week of course it was
impossible, but as it would have been equally impossible for them to
have bought anything it did not so much matter. The question was
how to get so large a sum of money round to the outlying post
offices in safety, for, though the city is now comparatively safe, there
are still snipers in outlying districts, and any party of Post Office
officials known to have possession of large sums of money would
undoubtedly have been attacked. So H. bethought him to requisition
for one of the boiler armoured cars with military guard, and it was at
once granted him. We had heard of them from N., but had not seen
one, and great was the excitement at the hotel when this huge
monster arrived for H.'s instructions. We all went out and examined
it.
It was not one of Guinness's, but one that had been rigged up by
one of the railway companies, with an engine boiler fixed on to a
huge motor trolley, all painted light grey; and all down each side
were black dots in an elegant design—something like this:—

Here and there one of these squares was cut out and acted as an
air-hole, but they all looked exactly alike, so a sniper on a roof or
from a window aiming at one of these squares probably found his
bullet struck iron and bounded off to the accompaniment of derisive
jeers from the "Tommies" inside.
 Armoured Car.

From the hotel the car proceeded to the Bank of Ireland, and took
over £10,000 in silver, and started on its round to all the post
offices, delivering the money in perfect safety. I will try and send you
a photograph of one of these most ingenious conveyances.
After it had started on its round I went with H. to see the
temporary sorting offices. H. had secured an enormous skating rink
at the back of the Rotunda, and here all the sorting of letters was
going on, with no apparatus whatever except what the men had
contrived for themselves out of seats, benches and old scenery.
They were all hard at work—a regular hive of bees. We think it is
greatly to the credit of the Post Office staff that in twelve days from
the outbreak of the rebellion and three days after the actual
cessation of hostilities the whole service was reorganised, with two
deliveries a day in Dublin, besides the ordinary country and mail
deliveries. The engineers and telegraphists were no less wonderful.
Indeed the staff from top to bottom of the office have worked
splendidly, and H. is very proud of them. We looked in at the poor
G.P.O. on our way back. It is still smouldering, and it will be quite a
fortnight before any excavations can be begun, but H. hopes to get
the safe that contains many of our treasures out of the wall and
opened in a few days.
To-day a Dr. C. who is staying in the hotel told me of an
extraordinary escape he had had during one of the days of the
rebellion. He was walking through one of the squares, which he had
been told was clear of snipers, with an old friend of about eighty,
when suddenly a bullet struck the pavement at the feet of his friend
and ricochetted off. It was within an inch of the old gentleman's feet,
and he was greatly interested, wanting to find the bullet to keep as
a memento. While they were looking about for it a man who had
been walking just behind them passed them on the pavement, and
had only gone a few yards when they heard a second rifle shot, and
the man dropped like a stone, shot through the heart. Dr. C. ran up
to him, but he was quite dead. There was absolutely no safety
anywhere from the snipers; man, woman, or child, nothing came
amiss to them. It was dastardly fighting, if it could be called fighting
at all.
A few days after St. Stephen's Green was supposed to have been
cleared of rebels, we were told of a young woman whose husband
was home from the war wounded and in one of the hospitals. She
was going to see him, so took a short cut through the Green, when
she was shot through the thigh; it is supposed by a rebel, in hiding
in the shrubberies.
Sunday, 7th.
I am sending off my other letters to you to-morrow, as we hear
the censorship is no longer so strict, and as from the papers the
position here seems now to be known in England private letters are
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

textbookfull.com