Text Categories and Corpus Users: A Response to David Lee
Aston, Guy, Language, Learning & Technology
In designing any corpus, it is necessary to decide what types of texts to include, and how many of each type. (I use the term "text type" as a neutral one which does not imply any specific theoretical stance.) The British National Corpus (Burnard, 1995) made an initial division into written texts and spoken ones (i.e. transcripts of recordings), and within each of these macrocategories, employed further categorisations and subcategorisations. For the spoken component, a first distinction was between "demographic" (conversations: 153 texts) versus "context-governed" (speech recorded in particular types of setting: 757 texts), and the "context-governed" component was further divided according to the nature of the setting (educational/informative; business; public/institutional; leisure: from 131 to 262 texts in each), paralleled by a monologue/dialogue distinction (40%/60%). For the written component, two principal parallel categorisations were used: "domain" (i.e., subject matter, divided into nine classes, viz., imaginative; arts; belief and thought; commerce; leisure; natural science; applied science; social science; world affairs; from 146 to 527 texts in each) and "medium" (five classes, viz., book; periodical; miscellaneous published; published; to be spoken; from 35 to 1,414 texts in each). All figures refer to the BNC World Edition (2001).
Text categorisations, as Lee notes, are generally based on "external" criteria -- where/when the text was produced, by/for who, what it is about -- rather than "internal" ones based on its linguistic characteristics. The categorisations used in corpus design tend to be broad rather than delicate, since what corpus designers want to do is to enable users to generalize about and compare different categories. To generalize with any confidence, each category must contain a substantial number of different texts, so that no one text exerts an undue influence on that category (early corpora such as Brown and Lob, which were relatively small, got around this problem by including very short samples from a large number of texts); and each category must contain a wide variety of different texts, so that no one subcategory exerts an undue influence on that category as a whole (Biber, 1993): the greater the variance within a category, the more texts will be needed in order to document that variance. Thus, it may be decided to include roughly equal numbers of texts from different parts of the country, by authors of different sexes/ages or from different types of settings. Within the BNC "context-governed" component, for instance, the "educational/informative" category was designed to include lectures, talks, classroom interaction, and news commentaries, drawing these from different types of institutions in different areas and with a wide range of speakers and topics.
Since corpora cannot be infinite, the delicacy of the categorisations to be employed is largely determined by practical considerations. The BNC, which contains just over 4,000 texts, uses a framework which guarantees at least 100 texts in most principal categories. You may or may not like the categories chosen, but the corpus arguably allows you to generalize about these categories -- about spoken and written texts, the nine different domains of written texts, the four different domains of "context-governed" spoken texts, and so forth -- with reasonable certainty that findings will not be unduly biased by any particular text or any particular subcategory of texts. These categories are indicated in the headers to individual texts as attributes of the [less than]catRef[greater than] element, using which it is possible to restrict queries to a particular category or combination of categories. A number of errors of categorisation in the first release of the BNC have been corrected in the World Edition (2001).
Users may, however, want to employ different categorisations from those employed by the corpus designers. …