Neural Representations of Natural Language

132 Pages·2019·4.024 MB·

by  Lyndon White, Roberto Togneri, Wei Liu, Mohammed Bennamoun

Studies in Computational Intelligence 783
Lyndon White · Roberto Togneri 
Wei Liu · Mohammed Bennamoun
Neural
Representations
of Natural
Language
Studies in Computational Intelligence
Volume 783
Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: [email protected]
The series “Studies in Computational Intelligence” (SCI) publishes new develop-
mentsandadvancesinthevariousareasofcomputationalintelligence—quicklyand
with a high quality. The intent is to cover the theory, applications, and design
methods of computational intelligence, as embedded in the fields of engineering,
computer science, physics and life sciences, as well as the methodologies behind
them. The series contains monographs, lecture notes and edited volumes in
computational intelligence spanning the areas of neural networks, connectionist
systems, genetic algorithms, evolutionary computation, artificial intelligence,
cellular automata, self-organizing systems, soft computing, fuzzy systems, and
hybrid intelligent systems. Of particular value to both the contributors and the
readership are the short publication timeframe and the world-wide distribution,
which enable both wide and rapid dissemination of research output.
More information about this series at http://www.springer.com/series/7092
Lyndon White (cid:129) Roberto Togneri
Wei Liu (cid:129) Mohammed Bennamoun
Neural Representations
of Natural Language
123
Lyndon White Wei Liu
Department ofElectrical, Electronic and Department ofComputer Scienceand
Computer Engineering, Schoolof Software Engineering, Schoolof Physics,
Engineering, Faculty of Engineeringand MathematicsandComputing, Faculty of
MathematicalSciences EngineeringandMathematical Sciences
TheUniversity of Western Australia TheUniversity of Western Australia
Perth, WA,Australia Perth, WA,Australia
RobertoTogneri Mohammed Bennamoun
Department ofElectrical, Electronic and Department ofComputer Scienceand
Computer Engineering, Schoolof Software Engineering, Schoolof Physics,
Engineering, Faculty of Engineeringand MathematicsandComputing, Faculty of
MathematicalSciences EngineeringandMathematical Sciences
TheUniversity of Western Australia TheUniversity of Western Australia
Perth, WA,Australia Perth, WA,Australia
ISSN 1860-949X ISSN 1860-9503 (electronic)
Studies in Computational Intelligence
ISBN978-981-13-0061-5 ISBN978-981-13-0062-2 (eBook)
https://doi.org/10.1007/978-981-13-0062-2
LibraryofCongressControlNumber:2018938363
©SpringerNatureSingaporePteLtd.2019
Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpart
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
orinformationstorageandretrieval,electronicadaptation,computersoftware,orbysimilarordissimilar
methodologynowknownorhereafterdeveloped.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publicationdoesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfrom
therelevantprotectivelawsandregulationsandthereforefreeforgeneraluse.
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
authorsortheeditorsgiveawarranty,expressorimplied,withrespecttothematerialcontainedhereinor
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictionalclaimsinpublishedmapsandinstitutionalaffiliations.
ThisSpringerimprintispublishedbytheregisteredcompanySpringerNatureSingaporePteLtd.
Theregisteredcompanyaddressis:152BeachRoad,#21-01/04GatewayEast,Singapore189721,
Singapore
Preface
Wedo,therefore,offeryouourpracticaljudgements
whereverwecan.Asyougainexperience,youwillformyour
ownopinionofhowreliableouradviceis.Beassuredthatit
isnotperfect!
Press,Teukolsky,VettingandFlannery,NumericalRecipes,
3rded.,2007
Some might wonder why one would be concerned with finding good representa-
tions for natural language. To answer that simply, all problem-solving is signifi-
cantly simplified by finding a good representation, whether that is reducing a
A wealth of other materials
WikipediaarticlesonMLandNLPtendtobeofreasonablequality,Coursera
offers several courses on the topics, Cross Validated Stack Exchange has
thousandsofQAs,andthedocumentationformanymachinelearninglibraries
often contains high-quality tutorials. Finally, preprints of the majority of the
papers in the fields are available on arXiv.
real-world problem into a system of inequalities to be solved by a constrained
optimizerorusingdomain-specificjargontocommunicatewithexpertsinthearea.
Thereisatrulyimmenseamountofexistinginformation,designedforconsumption
by humans. Much of it is in text form. People prefer to create and consume
information in natural language (such as prose) format, rather than in some more
computationallyaccessibleformat(suchasadirectedgraph).Forexample,doctors
prefer to dictate clinical notes over filling in forms for a database.
One of the core attractions of a deep learning system is that it functions by
learning increasingly abstract representations of inputs. These abstract representa-
tions include useful features for the task. Importantly, they also include features
applicable even for more distantly related tasks.
Itisthegeneralassumptionofthisbookthatitisnotbeingreadinisolation.That
thereader isnotbereft ofallothersourcesofknowledge.Weassumenotonlycan
v
vi Preface
the reader access the primary sources for the papers we cite, but also that they are
abletodiscoverandaccessothersuitablematerials,inordertogodeeperonrelated
areas.Thereexistawealthofblogposts,videotutorials,encyclopaedicarticle,etc.,
on machine learning and on the mathematics involved.
In general, we do assume the reader is familiar with matrix multiplication. In
general, we will define networks in terms of matrices (rather than sums), as this is
more concise, and better reflects real-world applications, in terms of code that a
programmerwouldwrite.Wealsoassumeabasicknowledgeofprobability,andan
evenmorebasicknowledgeofEnglishlinguistics.Thoughtheyarenottheintended
audience, very little of the book should be beyond someone with a high-school
education.
The core focus of this book is to summarize the techniques that have been
developed over the past 1–2 decades. In doing so, we detail the works of several
dozen papers (and cite well over a hundred). Significant effort has gone into
describingtheseworksclearlyandconsistently.Indoingso,thenotationuseddoes
notnormallyline up with theoriginal notations usedinthose papers;however, we
ensureitisalwaysmathematicallyequivalent.Forsometechniques,theexplanation
isshortandsimple.Forothermorechallengingideas,ourexplanationmaybemuch
longer than the original very concise formulation that is common in some papers.
Forbrevity,wehavehadtolimitdiscussionofsomeaspectofnaturallanguage.
In particular, we have neglected all discussion on the notion that a word may be
made up of multiple tokens, for example, made up of. Phrases do receive some
discussion in Chap. 6. Works such as Yin and Schütze (2014) exploration deserve
more attention than we have space to give them.
Similarly, we do not have space to cover character-based models, mainly
characterRNNs,andotherdeepcharactermodelssuchasZhangandLeCun(2015).
Thesemodelshaverelevancebothassourcesfromwhichwordrepresentationscan
bederived(Bojanowskietal.2017),butmoregenerallycanbeusedforend-to-end
systems. Using a purely character-based approach forces the model to learn tok-
enizing, parsing, and any other feature extraction internally. We focus instead on
models that work within larger systems that accomplish such preprocessing.
Finally,we omitall discussion of attention models in recurrent neural networks
(Bahdanau et al. 2014). The principles to the application remain similar to using
normal recurrent neural networks in the applications discussed in this book. The
additional attention featuresallowseveralimprovements,makingitthestateofthe
art for many sequential tasks. We trust the reader with the experience of such
modelsandwillseehowthey maybeappliedtotheuses ofRNNsdiscussedhere.
This book makes extensive use of text in grey boxes. This is used to provide
remindersofdefinitionsandcommentsonpotentiallyunfamiliarnotations.Aswell
as to highlight non-technical aspects (such as societal concerns), and conversely
overly technical aspects (such as implementation details) of the works discussed.
They arealso usedtogive thetitlestocitations(to save flickingtothehugelist of
references at the back of the book) and for the captions to the figures and more
Preface vii
generally to provide non-essential but potentially helpful information without
interruptingtheflowofthemaintext.Onecouldreadthewholebookwithoutever
reading any of the text in grey boxes, as they are not required to understand the
main text. However, one would be missing out on a portion of, some very inter-
esting, content.
Perth, Australia Lyndon White
Roberto Togneri
Wei Liu
Mohammed Bennamoun
References
Bahdanau,Dzmitry,KyunghyunCho,andYoshuaBengio.2014.Neuralmachinetranslationby
jointlylearningtoalignandtranslate.InCoRRabs/1409.0473.arXiv:1409.0473.
Bojanowski, Piotr, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word
vectors with subword information. Transactions of the Association for Computational
Linguistics,vol.5,135–146.
Yin,WenpengandHinrichSchütze.2014.Anexplorationofembeddingsforgeneralizedphrases.
InACL2014,p.41.
Zhang, Xiang and Yann Le Cun. 2015. Text understanding from scratch. In CoRR proceedings
ofthe52ndannualmeetingoftheassociationforcomputationallinguistics,vol.11.
Contents
1 Introduction to Neural Networks for Machine Learning . . . . . . . . . 1
1.1 Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Function Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Network Hyper-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.1 Width and Depth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Activation Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Training. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4.1 Gradient Descent and Back-Propagation . . . . . . . . . . . . . . 12
1.5 Some Examples of Common Neural Network Architectures . . . . . 14
1.5.1 Classifier. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5.2 Bottlenecking Autoencoder. . . . . . . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2 Recurrent Neural Networks for Sequential Processing . . . . . . . . . . . 23
2.1 Recurrent Neural Networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2 General RNN Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.1 Matched-Sequence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.2 Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.3 Decoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2.4 Encoder-Decoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Inside the Recurrent Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.1 Basic Recurrent Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.2 Gated Recurrent Unit. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.3 LSTM Recurrent Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Further Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.1 Deep Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.2 Bidirectional RNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.3 Other RNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
ix
x Contents
3 Word Representations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1 Representations for Language Modeling. . . . . . . . . . . . . . . . . . . . 39
3.1.1 The Neural Probabilistic Language Model. . . . . . . . . . . . . 41
3.1.2 RNN Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Acausal Language Modeling. . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2.1 Continuous Bag of Words . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2.2 Skip-Gram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2.3 Analogy Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3 Co-location Factorisation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.1 GloVe. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.2 Further Equivalence of Co-location Prediction to
Factorisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Hierarchical Softmax and Negative Sampling. . . . . . . . . . . . . . . . 56
3.4.1 Hierarchical Softmax . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4.2 Negative Sampling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.5 Natural Language Applications – Beyond Language
Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.5.1 Using Word Embeddings as Features . . . . . . . . . . . . . . . . 67
3.6 Aligning Vector Spaces Across Languages. . . . . . . . . . . . . . . . . . 67
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4 Word Sense Representations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.1 Word Senses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.1.1 Word Sense Disambiguation. . . . . . . . . . . . . . . . . . . . . . . 75
4.2 Word Sense Representation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2.1 Directly Supervised Method. . . . . . . . . . . . . . . . . . . . . . . 79
4.2.2 Word Embedding-Based Disambiguation Method . . . . . . . 79
4.3 Word Sense Induction (WSI) . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.3.1 Context Clustering-Based Approaches. . . . . . . . . . . . . . . . 83
4.3.2 Co-location Prediction-Based Approaches . . . . . . . . . . . . . 86
4.4 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5 Sentence Representations and Beyond . . . . . . . . . . . . . . . . . . . . . . . 93
5.1 Unordered and Weakly Ordered Representations . . . . . . . . . . . . . 94
5.1.1 Sums of Word Embeddings . . . . . . . . . . . . . . . . . . . . . . . 95
5.1.2 Paragraph Vector Models (Defunct) . . . . . . . . . . . . . . . . . 96
5.2 Sequential Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2.1 VAE and Encoder-Decoder . . . . . . . . . . . . . . . . . . . . . . . 98
5.2.2 Skip-Thought . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99