Post on 27-Dec-2015
BTANT 129 w5
Introduction to corpus linguistics
BTANT 129 w5
Corpus
• The old school concept– A collection of texts especially if complete and
self-contained: the corpus of Anglo-Saxon verse
The Oxford Companion to the English Language
• The modern view– A collection of naturally occurring language
text chosen to characterize a state or variety of a language
• John Sinclair Corpus Concordance Collocation OUP
BTANT 129 w5
Corpus vs. archive
• Text archive• Collection of texts in their original format(Oxford Text Archive:
http://ota.ox.ac.uk/)• Corpus• texts collected and processed in a unified,
systematic mannerBritish National Corpus:
http://www.natcorp.ox.ac.uk/
BTANT 129 w5
BTANT 129 w5
BTANT 129 w5
Short history
Brief mention of just a select few! • Brown Corpus (Brown university)
– 1 m words– 15 genres– 500 samples 2000 words each– Area: US– Time: 1961
• LOB Corpus (Lancaster-Bergen-Oslo)– GB replica of Brown
BTANT 129 w5
Cobuild
• Major corpus initiative by Collins and Birmingham Univ. John Sinclair
• 1991 20 m • -> Bank of English currently 450 m
words• http://www.cobuild.collins.co.uk
BTANT 129 w5
British National Corpus
• 100 m words careful selection• 10 % spoken material• time span 1960 (fiction) – 1975 non-
ficion)• 40-50 000 word texts• TEI compliant SGML coding• http://www.comp.lancs.ac.uk/ucrel/
bncindex/
BTANT 129 w5
BTANT 129 w5
International Corpus of English
• 20 corpora of 1 m words devoted to varieties of English around the world
• 500 texts (300 written 200 spoken) of 2000 words each
• time span: 1990-0996• ICE-GB available in demo version• syntactic annotation, graphical tool
ICECUP
BTANT 129 w5
BTANT 129 w5
Corpus processing: tokenization
• Preprocessing– tokenization segmenting the text into
sentences• sometimes tricky: sentence delimiters in
mid-sentence positions
words• multi-word units – problem
– Normalization• restoring clitics, abbreviations ("can't",
"I've")
BTANT 129 w5
Corpus processing: tagging
• Tagging– labelling every word with its Part of
Speech category– Problem: ambiguity
• out of context, words can belong to different part of speech or have different analysis within the same POS
– set N vs. set V– bánt 'bánik' VBD vagy 'bánt' VBZ
BTANT 129 w5
Corpus processing: disambiguation
• Disambiguation– defining the correct analysis in context
• Two approaches:• both needs manually corrected training
corpus– statistical
• Hidden Markov model• calculating probability within a span of usually one or
two words• rate of success can be around 98%
– rule-based
BTANT 129 w5
Syntactic annotation
• Difficult to do on such a scale • shallow parsing• Treebank:
collection of syntactically analyzed sentences
• Penn treebank• http://www.cis.upenn.edu/~treebank/
BTANT 129 w5
Recent trends
• Word sense ambiguation (SENSEVAL) • http://www.itri.brighton.ac.uk/events/
senseval/
• Message understanding• http://www.itl.nist.gov/iaui/894.02/related_
projects/muc/index.html
• SEMANTIC WEB• making information on the web
understandable for machines• a vision requiring a huge effort, not clear
whether feasible at all
BTANT 129 w5
Representative sample?
• A corpus any size is inevitably a sample
• Of what?• Two approaches
– sampling speakers – demographic sampling
– sampling their output – text type sample
BTANT 129 w5
The notion of representativeness
• Sample vs. population• sample should be proportional to the
population for a given feature– example for demographic samplingif we know from census figures that 48% of
people in living in Budapest are malewe should compile our sample so that 48% of the
informants are male-> our sample is representative of Budapest
residents for gender
BTANT 129 w5
Trouble with representativeness
• What should be the units of sampling?• Registers, text types, genres etc.• But no independent evidence about
theirratio in the totality of language output
-> representativeness is an ideal but impossible to implement
BTANT 129 w5
Approaches to Representativeness
• Douglas Biber:• Rejects notion of proportional
sampling• Sample should be as varied as
possible• Representativeness measured in
terms of wide variety of text types included in the sample
BTANT 129 w5
The Web as a corpus?
• Pro:• immense database• dynamically
growing• ideal 'quick and
dirty' method
• Cons:• lots of rubbish,
irrelevant data• difficult to extract
hits• no language analysis• only string query,
which is crude
BTANT 129 w5
One quick example
• Representativity or representativeness
• Throw the two words at Google and have a look at the figures
• Think about the conclusions• There are special front-end sites
BTANT 129 w5
BTANT 129 w5
BTANT 129 w5
BTANT 129 w5