Corpus Linguistics and the Automatic Analysis of English

In a recent paper advocating a corpus-based and probabilistic approach to grammar development, Black, Lafferty, and Roukos (1992) argue that "the current state of the art is far from being able to produce a robust parser of general English" and advocate "steady and quantifiable," empirically corpus-driven grammar development and testing. Black et al. are addressing a community in which armchair introspection has been and still is the dominant methodology in many quarters, but in some parts of Europe, corpus linguistics never died. For nearly two decades, the Nijmegen group led by Jan Aarts have been undertaking corpus analyses that, although motivated primarily by the desire to study language variation using corpus data, are particularly relevant to the issue of broad-coverage grammar development. In distinction to other groups undertaking corpus-based work (e.g., Garside, Leech, and Sampson 1987), the Nijmegen group has consistently adopted the position that it is possible and desirable to develop a formal, generative grammar that characterizes the syntactic properties of a given corpus and can be used to assign appropriate analyses to each of its sentences. Nelleke Oostdijk's book provides a detailed description of the cumulative development of a grammar capable of analyzing a one million-word corpus of English written texts, drawn from a wide but balanced variety of sources. This task forms a significant component of the wider Tools for Syntactic Corpus Analysis (TOSCA) project being undertaken at Nijmegen. Oostdijk's work provides an excellent example of the strengths and weaknesses of the approach advocated by Black et al. In addition, she discusses issues such as sampling and tokenization of corpus material, as well as the exploitation of the analyzed corpus in studies of language variation. However, in this review I will concentrate on the central core of her book: the development of the grammar and performance of the associated parser, since this is the part that is most relevant to computational linguistics. Oostdijk begins by locating her work and the TOSCA project within the field of computational linguistics (arguing that it is distinguished by "an interest in language itself as it is actually produced" (p. 2)) and contrasting it to the LSP system (Sager 1981) and Parsifal (Marcus 1980). The comparison is brief and the choice odd since more general broad-coverage grammars, such as DIAGRAM (Robinson 1982), PEG (Jensen et al. 1986) and ANLT (Grover et al. 1989), and more corpus-oriented parsing systems, such as FIDDITCH (Hindle 1983, 1993) or MITFP (de Marcken 1990), have been developed within the field, but are not discussed anywhere. A similar suspicion of isolationism recurs in the sections dealing with the grammatical formalism used;