Hybrid Processing for Grammar and Style Checking

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 153–160 Manchester, August 2008 Hybrid processing for grammar and style checking Berthold Crysmann♣, Nuria Bertomeu♠, Peter Adolphs♥, Dan Flickinger♦, Tina Kl¨uwer♣♥ ♣ Universit¨at Bonn, Poppeldorfer Allee 47, D-53115 Bonn, {bcr,tkl}@ifk.uni-bonn.de ♠ Zentrum f¨ur Allgemeine Sprachwissenschaft, Berlin, nuria.bertomeu@dfki.de ♥ DFKI GmbH, Berlin, {peter.adolphs,kluewer}@dfki.de ♦ CSLI, Stanford University, danf@csli.stanford.edu Abstract This paper presents an implemented hy- brid approach to grammar and style checking, combining an industrial pattern- based grammar and style checker with bi- directional, large-scale HPSG grammars for German and English. Under this ap- proach, deep processing is applied selec- tively based on the error hypotheses of a shallow system. We have conducted a com- parative evaluation of the two components, supporting an integration scenario where the shallow system is best used for error de- tection, whereas the HPSG grammars add error correction for both grammar and con- trolled language style errors. 1 Introduction With the enormous amount of multilingual techni- cal documentation produced by companies nowa- days grammar and controlled language checking (henceforth: style checking) is becoming an appli- cation highly in demand. It is not only a helpful tool for authors, but also facilitates the translation of documents into foreign languages. Through the use of controlled language by the authors, docu- ments can be automatically translated more suc- cessfully than with the use of free language. Style checking should make authors aware of the con- structions which should not be used, as well as aiding in reformulating them. This can save a lot of translation costs for companies producing large amounts of mulitilingual documentation. Another application of grammar and style checking is the development of tutorial systems for learning a for- eign language, as well as any kind of authoring sys- tem for non-native speakers. Previous approaches to grammar and style checking can be divided into those based on fi- nite state methods and those based on linguisti- cally motivated grammars. To the former group be- long e.g. the systems FLAG (Bredenkamp et al., 2000a; Bredenkamp et al., 2000b) and MultiLint c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. (Haller, 1996; Schmidt-Wigger, 1998). The basic approach taken by such systems is the description of error patterns through finite state automata. The automata access the textual input enriched with annotations from shallow linguistic analysis com- ponents, such as part-of-speech tagging, morphol- ogy and chunking. In FLAG, for instance, the an- notation delivered by the shallow components is integrated into a complex feature structure. Rules are defined as finite state automata over feature structures. The great advantages of such systems are their robustness and efficient processing, which make them highly suitable for real-life grammar and style checking applications. However, since shallow modules usually cannot provide a full syn- tactic analysis, the coverage of these systems is limited to error types not requiring a broader (non- local) syntactic context for their detection. There- fore their precision in the recognition of non-local …

Content truncated to 500 words