Platform for Full-Syntax Grammar Development Using Meta-grammar Constructs

Platform for Full-Syntax Grammar Development Using Meta-grammar Constructs Aleˇs Hor´ak and Vladim´ır Kadlec Faculty of Informatics, Masaryk University Brno Botanick´a 68a, 602 00 Brno, Czech Republic {hales,xkadlec}@fi.muni.cz Abstract. This paper describes a combination of tools necessary for full or deep syntactic parsing of natural language – the syntactic parser synt, the graphical Grammar Development Workbench, GDW and the VerbaLex verb valency lexi- con tools. We describe the development of the mentioned tools and how they integrate into one system that allows a team of experts (computational linguists as well as pro- grammers) to cooperate on the development of grammar covering all Czech lan- guage phenomena. 1 Introduction Full parsing of natural language sentences is a complicated task that provides a tran- sition from unstructured text into structural information suitable for all forms of in- formation retrieval. Any kind of higher level language understanding and/or semantic processing must rely on the results of syntactic parsing. The quality of state-of-the-art syntactic parsers [1, 2] is still not completely satisfiable even for analytical languages like English. In case of free-word-order languages like Czech, the situation is even more complicated. At the Centre for Natural Language Processing at the Faculty of Informatics, Masa- ryk University in Brno, a full syntactic parsing system is being developed since 1997. The core of the system, the syntactic parser synt is based on the meta-grammar for- malism (e.g. [3–5]), which allows to specify complex sentence constituent combination rules in a maintainable way. The composition of the formalism resembles structures in Lexical functional grammars [6] – the meta-grammar rules are expanded within the guidance of combinatorial constructs plus contextual constraints and are supplemented with additional actions and agreement tests. The described full grammar development platform consists of the following com- ponents: – the Czech morphological analyser ajka [7]. This tagger provides non-disambi- guated output of plain text words. The ajka tagger covers almost 400.000 Czech word lemmas, which generates over 6 million word forms. – the VerbaLex verb valency lexicon tools [8]. The contextual constraints that safe- guard the syntactic analysis of free-word order language need extra lexical-semantic 311 2 A.Hor´ak, V.Kadlec Fig. 1. Interaction of our tools information about the sentence constituents. The creation of new verb valency lexi- con which contains so called complex valency frames has thus become a part of the grammar development platform. – the deep syntactic parser synt, see [9–11]. The parser development concentrates on high coverage on general corpus sentences (> 90 %). The parser is able to work with several parsing algorithms (GLR parsing, top-down, bottom-up and head- driven chart parsing). The synt parser also introduced a new variant of the head- driven technique, the head-driven dependent dot move [12] algorithm. This parsing technique allows to parse natural language sentences within median time of less than 0.1s/sentence. – the user interface for linguists – the Grammar Development Workbench, GDW [13]. The parser synt provides a command line interface, which is suitable for all forms of batch processing. However, the grammar development …

Content truncated to 500 words