Corpus-based methods and hand-built methods

Recent success of statistical corpus-based methods in a variety of areas of speech and language processing has led to the widespread view that traditional hand-built “rule-based” approaches are moribund. This is a misconception. As I shall argue in this talk, it is unlikely that rule-based approaches will ever be eliminated. Two examples are given to support this conclusion; one where the linguistic facts, though highly complex are basically quite regular; and another where the linguistic fact is exceedingly simple (hence hardly worth the effort of inferring from data), but where adding in this information can improve the output of a statistical model.