Theory-driven and Corpus-driven Computational Linguistics and the Use of Corpora

Computational linguistics and corpus linguistics are closely-related disciplines: they both exploit electronic corpora, extract various kinds of linguistic information from them, and make use of the same methods to acquire this information. Moreover, both were heavily affected by "paradigm shifts" from the prevailing empiricism of the 1950s, to rationalism, then back again with a revival of empirical methods in the 1990s. Computational linguistics deals with the formal modeling of natural language. The formal models can be used to draw conclusions about the structure and functioning of the human language system. They also form the basis of implemented systems for the analysis and generation of spoken or written language, in a variety of applications. The methods applied in building these models are of different kinds since, as a result of the above-mentioned paradigm changes, work in computational linguistics has taken two different paths. Both branches of computational linguistics aim to build models of natural language, but each exploits different techniques: the rationalist's branch focuses on theory-driven, symbolic, nonstatistical methods, whilst the empiricist's branch focuses on corpus-driven and statistical techniques. As we will see later however, the distinction between the branches is these days less clear, and the two fields seem to be coming together again as people successfully combine concepts and methods from each field. Obviously, the corpus-driven branch of computational linguistics has a natural affinity to corpus linguistics, and a shared interest in corpus exploitation. As a consequence, many research topics can be attributed equally well to either computational linguistics or corpus linguistics; examples include part-of-speech tagging (see article 25), treebanking (article 17), semantic tagging (article 27), coreference resolution (article 28), to name just a few. At opposite extremes of computational and corpus linguistics, the ultimate goals of corpus exploitation do however diverge: certain domains of corpus-driven computational linguistics aim to build "optimal" models "no matter how", and the particular corpus features that find their way into such models are not seen as interesting per se; in contrast, corpus linguistics could be said to target exactly these features, the "ingredients" of the models.