Natural Language Processing for Corpus Linguistics

96 Pages·2022·5.155 MB·

by  Jonathan Dunn

D
Corpus analysis can be expanded and scaled up by u
n
incorporating computational methods from natural language n
processing. This Element shows how text classification and text
similarity models can extend our ability to undertake corpus
linguistics across very large corpora. These computational
methods are becoming increasingly important as corpora Corpus Linguistics
grow too large for more traditional types of linguistic analysis.
We draw on five case studies to show how and why to use
computational methods, ranging from usage-based grammar
to authorship analysis to using social media for corpus-based
sociolinguistics. Each section is accompanied by an interactive
code notebook that shows how to implement the analysis in
n natural Language
Python. A stand-alone Python package is also available to help a
t
readers use these methods with their own data. Because large- u
r
a
scale analysis introduces new ethical problems, this Element l L Processing for
pairs each new methodology with a discussion of potential an
g
ethical implications. u
a Corpus Linguistics
g
e
P
r
o
c
e
About the Series Series Editor s
s
Corpus Linguistics has grown to become Susan Hunston in
g
part of the mainstream of Linguistics and University of f
o
Applied Linguistics. This Elements series Birmingham r C Jonathan Dunn
is designed to meet the needs of students o
r
p
and researchers who need to keep up with u
s
this changing field, including introductions L
in
to main topics areas as well as accounts
g
u
of the latest ideas and developments. is
t
ic
s
Cover image: monsitj / iStock / Getty Images Plus Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to thIeS SCNam 26b3r2id-8g0e9 C7o (roen ltienrem)s
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447 ISSN 2632-8089 (print)
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
Elements in Corpus Linguistics
editedby
SusanHunston
UniversityofBirmingham
NATURAL LANGUAGE
PROCESSING FOR CORPUS
LINGUISTICS
Jonathan Dunn
University of Canterbury
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
UniversityPrintingHouse,CambridgeCB28BS,UnitedKingdom
OneLibertyPlaza,20thFloor,NewYork,NY10006,USA
477WilliamstownRoad,PortMelbourne,VIC3207,Australia
314–321,3rdFloor,Plot3,SplendorForum,JasolaDistrictCentre,
NewDelhi–110025,India
103PenangRoad,#05–06/07,VisioncrestCommercial,Singapore238467
CambridgeUniversityPressispartoftheUniversityofCambridge.
ItfurtherstheUniversity’smissionbydisseminatingknowledgeinthepursuitof
education,learning,andresearchatthehighestinternationallevelsofexcellence.
www.cambridge.org
Informationonthistitle:www.cambridge.org/9781009074438
DOI:10.1017/9781009070447
©JonathanDunn2022
Thispublicationisincopyright.Subjecttostatutoryexception
andtotheprovisionsofrelevantcollectivelicensingagreements,
noreproductionofanypartmaytakeplacewithoutthewritten
permissionofCambridgeUniversityPress.
Firstpublished2022
AcataloguerecordforthispublicationisavailablefromtheBritishLibrary.
ISBN978-1-009-07443-8Paperback
ISSN2632-8097(online)
ISSN2632-8089(print)
Additionalresourcesforthispublicationatwww.cambridge.org/dunnresources
CambridgeUniversityPresshasnoresponsibilityforthepersistenceoraccuracyof
URLsforexternalorthird-partyinternetwebsitesreferredtointhispublication
anddoesnotguaranteethatanycontentonsuchwebsitesis,orwillremain,
accurateorappropriate.
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
Natural Language Processing for Corpus Linguistics
ElementsinCorpusLinguistics
DOI:10.1017/9781009070447
Firstpublishedonline:March2022
JonathanDunn
UniversityofCanterbury
Authorforcorrespondence:JonathanDunn,
[email protected]
Abstract:Corpusanalysiscanbeexpandedandscaledupby
incorporatingcomputationalmethodsfromnaturallanguageprocessing.
ThisElementshowshowtextclassificationandtextsimilaritymodelscan
extendourabilitytoundertakecorpuslinguisticsacrossverylargecorpora.
Thesecomputationalmethodsarebecomingincreasinglyimportantas
corporagrowtoolargeformoretraditionaltypesoflinguisticanalysis.We
drawonfivecasestudiestoshowhowandwhytousecomputational
methods,rangingfromusage-basedgrammartoauthorshipanalysisto
usingsocialmediaforcorpus-basedsociolinguistics.Eachsectionis
accompaniedbyaninteractivecodenotebookthatshowshowto
implementtheanalysisinPython.Astand-alonePythonpackageisalso
availabletohelpreadersusethesemethodswiththeirowndata.Because
large-scaleanalysisintroducesnewethicalproblems,thisElementpairs
eachnewmethodologywithadiscussionofpotentialethicalimplications.
ThisElementalsohasavideoabstract:www.cambridge.org/dunnabstract
Keywords:computationallinguistics,naturallanguageprocessing,corpus
linguistics,textclassification,textsimilarity,usage-basedgrammar,
corpus-basedsociolinguistics,computationalstylistics,computationalsyntax
JELclassifications:A12,B34,C56,D78,E90
©JonathanDunn2022
ISBNs:9781009074438(PB),9781009070447(OC)
ISSNs:2632-8097(online),2632-8089(print)
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
Contents
1 Computational Linguistic Analysis 1
2 Text Classification 13
3 Text Similarity 39
4 Validation and Visualization 62
5 Conclusions 79
References 81
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
Accessing the Code Notebooks
https://doi.org/10.24433/CO.3402613.v1
https://github.com/jonathandunn/text_analytics
https://github.com/jonathandunn/corpus_analysis
TorunthenotebooksthroughCodeOcean,youwillneedtoclickthecommand
thatsays“EditYourCopy”inthetopright-handcorner,asshowninthefirst
screenshot:
The“Jupyter”commandwillnowbeavailableundertheheading“Reproduci-
bleRun”asshowninthesecondscreenshot:
This will start up the interactive notebook container. You can now find the
notebookswithinthe“code”folder.
Thefollowingisalistofinteractivenotebookstogetherwiththesectionofthe
Elementwhichtheyaccompany:
Lab1.2.AccessingtheCorpora
Lab1.3.VisualizingCategories
Lab1.4.UsingGroupbytoExploreCategories
Lab1.5.VectorizingTexts
Lab2.1.GettingxandyArraysforDialects
Lab2.2.ClassifyingCitieswithTF-IDFandPMI
Lab2.3.ClassifyingAuthorswithFunctionWordN-Grams
Lab2.4.UsingPositionalVectorsforPartsofSpeech
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
Lab2.5.ClassifyingHotelsbyQualityUsingSentimentAnalysis
Lab2.7.ClassifyingCitiesUsingMLPs
Lab3.2.RegisterandCorpusSimilarity
Lab3.3.FindingSimilarDocuments
Lab3.4.FindingAssociatedWords
Lab3.5.WorkingwithWordEmbeddings
Lab3.6.ClusteringWordEmbeddings
Lab4.1.BaselinesforClassifyingPoliticalSpeeches
Lab4.2.EnsuringValidityUsingCross-Validation
Lab4.3.UnmaskingAuthorship
Lab4.4.ComparingWordEmbeddings
Lab4.5.MakingMapsforLinguisticDiversity
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
NaturalLanguageProcessingforCorpusLinguistics 1
1 Computational Linguistic Analysis
1.1 ScalingUp Corpus Linguistics
Corpuslinguisticshasenteredagoldenage,drivenbyboththeamountandthe
rangeoflanguagethatisnowavailableforlinguisticanalysis.Corpusdatais
abletorepresentapopulation’susageatscale,bypassingthelimitationswhich
made introspection so important in the 1950s. But this wide availability of
languagedatarequiresthatlinguistshavethemethodsavailabletoanalyzeit.
Andwhiletherehasbeenasurgeofadvancesinnaturallanguageprocessing
and computational linguistics, these advances have become increasingly dis-
connected from corpus linguistics and linguistic theory. This Element brings
naturallanguageprocessingandcorpuslinguisticstogether,showinghowcom-
putationalmodelscanbeusedtoanswerbothcategorizationandcomparison
problems. These computational models are presented using five case studies
thatwillbeintroducedinthenextsection,rangingfromsyntacticanalysisto
registeranalysistocorpus-basedsociolinguistics.
The goal here is to show how to use these computational models, what
linguisticquestionstheycananswer,andwhyitisimportanttoscaleupcorpus
linguisticsinthisway.AlinguistcanusethisElementtolearnhowtousenatu-
rallanguageprocessingtoanswerlinguisticquestionstheyarealreadyfamiliar
with.AndacomputerscientistcanusethisElementtolearnaboutthelinguis-
ticassumptionsandlimitationsbehindcomputationalmethods,mattersthatare
toooftendisregardedwithinnaturallanguageprocessingitself.
Acategorizationproblemisaboutassigningapredefinedlabeltosomepiece
oflanguage.Atthewordlevel,thiscouldinvolveaskingwhetheraparticular
open-classwordisanounoraverb.Atthesentencelevel,thiscouldbeasking
what kind of construction a particular sentence represents. At the document
level,thiscouldbeaskingwhetheraparticularspeakerrepresentsNewZealand
EnglishorAustralianEnglish.Allofthesequestionscanbeansweredusinga
text classifier. This is a type of supervised machine learning in which we as
linguistsdefinethecategoriesthatweareinterestedin.
A comparison problem is about measuring the relationship between two
observations.Atthewordlevel,thiscouldbeaskingwhethertwonounslike
cat and dog belong to the same semantic domain. At the sentence level,
this could be asking whether two tweets have a similar sentiment. At the
document level, this could be asking whether two articles are examples of
a similar style. These questions can be approached using a text similarity
model. This is a type of unsupervised machine learning in which we as lin-
guistsonlycontroltherepresentationsbeingused,notthesetoflabelsusedfor
annotation.
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447
2 CorpusLinguistics
Howwelldocomputationalmodelscomparewithhumanintrospection?In
some cases, models can reproduce human intuitions with a high degree of
accuracy. For example, text classifiers have been shown to make very good
predictions about the part of speech of individual words when trained on
small amounts of annotated data. In a case like this, a small amount of seed
data, which is annotated by a linguist, supports the analysis of corpora too
large to be annotated by a linguist. So a text classifier allows us to scale up
introspection-basedannotations.
Inothercases,computationalmodelscandetectpatternsinlanguagethatare
not visible to human introspection. For example, research in both authorship
analysisanddialectidentificationhasshownthatthereareenoughindividual-
specificandcommunity-specificvariantstoenableaccuratepredictionsofwho
producedaspecificdocument.But,aslinguists,ourownintrospectionsarenot
preciseenoughtoidentifythesesamepatterns.Inacaselikethis,computational
linguisticanalysismakesitpossibletoanswernewquestionsaboutlanguage.
Finally,therearecaseswherecomputationalmodelscompletelymisssome-
thing that is easily accessible to humans. For example, we will follow a case
study on multilingualism online which shows that 90 percent of digital lan-
guagedata(fromthewebandsocialmedia)representsjusttwentylanguages.
Mostlanguagesintheworldarelow-resourcelanguagesfromacomputational
perspective.Asaresult,manyofthecomputationalmethodsthatwecoverin
this Element are difficult to apply to these languages. As linguists, however,
we do not require millions or billions of words in a language before we can
beginouranalysis.
Weneedcomputationallinguisticanalysisfortworeasons:for reproduci-
bilityandforscalability.First,everystepinacomputationalpipelineisfully
automated,whichmeansthatitcanbereproducedandverified.Forexample,
this Element follows five separate case studies that we will introduce in the
nextsubsection.AllthegraphsandfiguresandexperimentsintheElementcan
bereproducedusingthecodenotebooksthatarelinkedwithineachsection.1
Thisisanexampleofhowcomputationalmethodssupportreproducibility.
Second, the once-revolutionary Brown Corpus contained 1 million words
(Francis & Kucera, 1967). But it is common now for corpora to range from
1billionwords,liketheGeoWACfamilyofcorpora(Dunn&Adams,2020),
up to 400 billion words, like the Corpus of Global Language Use (Dunn,
2020).Theseverylargecorporaareoftendrawnfromdigitalsourceslikethe
web, social media, Wikipedia, and news articles. While these sources of lan-
guage data have tremendous potential for testing linguistic hypotheses on a
1 Andathttps://github.com/jonathandunn/corpus_analysis
Downloaded from https://www.cambridge.org/core. IP address: 179.113.62.209, on 08 Mar 2022 at 17:22:37, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781009070447