Principles of corpus construction Matthew Brook ODonnell University

Principles of corpus construction Matthew Brook O’Donnell University of Liverpool - Corpus Linguistics Summer Institute 2008

Aims l l l What is a corpus? What principles guide the construction, development and selection of a corpus? When and How to build a corpus Can the web be used for building corpora? Workshop: Build a small corpus of web texts

What is a corpus? l John Sinclair (1933 -2007)

What is a corpus? l John Sinclair (1933 -2007) A corpus is a collection of pieces of language text in electronic form, selected according to external criteria to represent, as far as possible, a language or language variety as a source of data for linguistic research. (Sinclair 2004)

Corpus l l l Authentic language data Electronic/machine readable form Designed and collected according to sampling procedures Representative of language For linguistic investigation

But first…. l l Let’s talk about food! How could we compile a representative list (=CORPUS) of food/dishes from around the world?

How can we group these foods? 1. Where they come from

1. Continental Food Corpus Europe Asia ? America (North & South)

How can we group these foods? 1. 2. Where they come from Their main component

2. Main component corpus Fish Meat Vegetarian

Where do you eat it?

How can we group these foods? 1. 2. 3. Where they come from Their main component Where you usually eat it

3. ‘Fast food potential’ corpus Takeaway Restaurant

How can we group these foods? 1. 2. 3. 4. Where they come from Their main component Where it is usually eaten What you use to eat it

4. Consumption Implement Corpus Knife & fork Chopsticks Hands

How can we group these foods? 1. 2. 3. 4. Where they come from Their main component Where it is usually eaten What you use to eat it

1+2. Continental & Main Component Corpus Fish Europe Asia America (North & South) Meat Veg.

Corpus l l l Authentic language data Electronic/machine readable form Designed and collected according to sampling procedures Representative of language For linguistic investigation

Language corpora l l l Different types of corpus Corpus size Sample size Representativeness - sampling Classification criteria

Types of corpus • Sample Corpus: a fixed sample of text, often used as a reference corpus for comparing • Monitor Corpus: a corpus which develops and is added to or filtered depending on the researcher’s needs • Mini-corpus: a small corpus (e. g. to be compared with a reference corpus) • Multilingual Corpus: corpus in a variety of languages

Types of corpus • Comparable Corpus: texts in 2 languages or 2 varieties but not matched up • Parallel Corpus: texts are translations of each other, eg. Canadian Hansard, corpus of versions of Plato, Bible • Translation Corpus: 2 or more sets of texts classified as either originals or translations, the purpose being to identify features of translation (Manchester: Baker) • Diachronic Corpus: Helsinki, LOB v. FLOB • Learner Corpus: texts are written by language learners

Corpus Size – Is bigger better? l 1 st Generation Corpora = 1 Million Words – – l 2 nd Generation Sample Corpora – l BNC, ANC = 100 Million Words Monitor Corpora – l BROWN, LOB ICE corpora Bank of English (450+ million and growing!) Specialized corpora – Depends on source and scope of problem under investigation

Sample Size ‘Personally I would like to see ‘whole text’ as a default condition, thus classifying sample corpora as one of the categories of special corpora. . . To me the use of small samples is just a remnant of the early restraints on corpus building, and the advantages of whole texts can be set out in powerful argument. The use of samples of constant size gains only a spurious air of scientific method, since it confers no benefit on the corpus, and is as practical as Genghis Khan’s fabled policy of having all his soldiers the same height. ’ (Sinclair 1995: 27 -28)

Sampling l Population – – Production Reception

Classifying Texts l Internal Criteria – – Topic (aboutness) Register/Style

Classifying Texts l Internal Criteria – – l Topic (aboutness) Register/Style External Criteria (situational parameters)

Brits treat English with such disdain By Mr M. Rasheed Iqbal Published: June 30 2007 03: 00 | Last updated: June 30 2007 03: 00 From Mr M. Rasheed Iqbal. Sir, I agree with Henry von Blumenthal (Letters, June 23). It is very discouraging to hear news presenters saying "gonna" and "wanna" on the BBC news. We were brought up to speak English properly and it is disappointing to see Brits treat the language with such disdain. I hope the BBC will stem the tide and pull up its socks. M. Rasheed Iqbal, National Bank of Dubai, Deira, Dubai, UAE Copyright The Financial Times Limited 2007

Mode l Primary Channel – Written l Format – Published (print & web) l Setting - Public

Tenor l Addressee – – l Plurality – individual (editor) / plural (readers) Presence – absent Interactiveness – written correspondence/response Shared knowledge – readers of same publication Addressor – – Demographic: Male, from Dubai, works in Bank, educated, at least bilingual? Acknowledgement: Self-identified in text

Field l Factuality – responding to actual event (TV broadcast), expressing personal opinion l Purposes – complain, express viewpoint, condemn slipping standards, correct perceived decline l Topics – use of British English on BBC, value of language education in former era

When and How to build a corpus 1. DON’T! – use one of the available corpora a. b. 2. Repurpose existing archive/collection a. 3. If interested in differences in conversational language in British English (age, sex, class etc. differences)… USE British National Corpus Combine and subsample existing corpora to match your Any electronic texts available – results of surveys, DA/CA transcripts Build your own! a. b. OCR, download, extract from PDF TYPE IT IN!!!!

Using the web as source for corpora

Web as corpus: Advantages l l l Massive (and expanding) amounts of electronic text Whole texts Wide reach of text-types/topics/genres Much in the public domain Google (etc) as corpus query tool

Web as corpus: Disadvantages l l So much text that balance is difficult to achieve Copyright is difficult to ascertain for many documents Pages containing large amounts of extraneous material (menus, formatting, graphics) Explosion of information… who is writing it? Who is reading it?