(Not my invention but a borrowing from Michael Stubbs)
What can corpora do and what they cannot do -
First, corpora do not provide negative evidence. This means that they cannot tell us what is possible or not possible. Everything included in a corpus is what language users have actually produced. A corpus, however large or balanced, cannot be exhaustive except in a very limited range of cases. Nevertheless, a representative corpus can show what is central and typical in language.
Second, corpora can yield findings but rarely provide explanations for what is observed. These explanations must be developed using other methodologies, including intuition.
Third, the use of corpora as a methodology also defines the boundaries of any given study. As we have emphasized throughout the book, the usefulness of corpora in language studies depends upon the research question being investigated. As Hunston (2002: 20) argues, ‘They are invaluable for doing what they do, and what they do not do must be done in another way.’ It is also important that readers learn how to formulate research questions amenable to corpus-based investigation.
Finally, it is important to keep in mind that the findings based on a particular corpus only tell us what is true in that corpus, though a representative corpus allows us to make reasonable generalizations about the population from which the corpus was sampled. Nevertheless, unwarranted generalizations can be misleading (see my posting on statistics in corpus lingusitics).
The development of the corpus-based approach as a tool in language studies has been compared to the invention of telescopes in astronomy (Stubbs 1996: 231). If it is ridiculous to criticize a telescope for not being a microscope, it is equally pointless to criticize the corpus-based approach for not doing what it is not intended to do (Stubbs 1999).
