Parsed Corpora for Linguistics

Proceedings of the EACL 2009 Workshop on the Interaction between Linguistics and Computational Linguistics, pages 33–39, Athens, Greece, 30 March, 2009. c⃝2009 Association for Computational Linguistics Parsed Corpora for Linguistics Gertjan van Noord University of Groningen G.J.M.van.noord@rug.nl Gosse Bouma University of Groningen G.Bouma@rug.nl Abstract Knowledge-based parsers are now accu- rate, fast and robust enough to be used to obtain syntactic annotations for very large corpora fully automatically. We argue that such parsed corpora are an interesting new resource for linguists. The argument is illustrated by means of a number of re- cent results which were established with the help of parsed corpora. 1 Introduction Once upon a time, knowledge-based parsers were slow, inaccurate and fragile. This is no longer true. In the last decade, enormous improvements have been achieved in this area. Parsers based on constraint-based formalisms such as HPSG, LFG, and CCG are now fast enough for many appli- cations; they are robust; and they perform much more accurately than previously by incorporat- ing, typically, a statistical disambiguation compo- nent. As a consequence, such parsers now obtain competitive, if not superior, performance. Zae- nen (2004), for instance, points out that the (LFG- based) XLE parser is fast, has a statistical disam- biguation component, and is robust, and thus al- lows full parsing to be incorporated in many appli- cations. Clark and Curran (2007) show that both accurate and highly efficient parsing is possible using a CCG. As a consequence of this development, massive amounts of parsed sentences now become avail- able. Such large collections of syntactically an- notated but not manually verified syntactic analy- ses are a very useful resource for many purposes. In this position paper we focus on one purpose: linguistic analysis. Our claim is, that very large parsed corpora are an important resource for lin- guists. Such very large parsed corpora can be used to search systematically for specific infre- quent syntactic configurations of interest, and also to obtain quantitative data about specific syntac- tic configurations. Although parsed corpora obvi- ously contain a certain amount of noise, for many applications the abundant size of these corpora compensates for this. In this paper, we illustrate our position by a nu- mer of recent linguistic studies in which very large corpora of Dutch have been employed, which were syntactically annotated by the freely avail- able Alpino parser (Bouma et al., 2001; van No- ord, 2006). The Alpino system incorporates a linguistically motivated, wide-coverage grammar for Dutch in the tradition of HPSG. It consists of over 800 grammar rules and a large lexicon of over 300,000 lexemes (including very many person names, geo- graphical names, and organization names) and var- ious rules to recognize special constructs such as named entities, temporal expressions, etc. Since we use Alpino to parse large amounts of data, it is crucial that the parser is capable to treat sen- tences with unknown words. A large set of heuris- tics have been implemented carefully to deal with unknown words and word sequences. Based on the …

Content truncated to 500 words