A Bare-bones Constraint Grammar

A Bare-bones Constraint Grammar Eckhard Bick Institute of Language and Communication, University of Southern Denmark Campusvej 55, DK 5230 Odense M eckhard.bick@mail.dk Abstract. This paper presents a solution for overcoming the lexical resource gap when mounting rule-based Constraint Grammar systems for minor languages, or in the face of licensing and financing limitations. We investigate how the performance of a CG disambiguation grammar responds to shifting input parameters, among them lexicon limitations of various degrees, the lack a morphological analyzer or both. We propose solutions for a bare-bones system, introducing endings heuristics and so-called morphological APPEND rules. For English, even with an unadapted disambiguation grammar, our bare-bones tagger achieved F-scores of 90-96% for part of speech, and 94- 97% for lemmatization, depending on the modules and mini-lexica used. Keywords: Less-resourced languages, PoS-tagging, Constraint Grammar, morphological analyzer, lexicon generation 1 Introduction Conventional wisdom has it that in the realm of natural language parsing (NLP), statistical methods are more cost-efficient and easier to build than rule-based systems. However, the latter are dependent on training data, and machine learning of morphosyntactic analysis relies on the existence of a fair-sized annotated corpus for the language in question. If no such corpus exists, manual annotation of a boot-strapping corpus will be necessary, eliminating part of the cost- effectiveness advantage. Furthermore, in the face of Zipf's law, the limited availability of large linguist-revised corpora (especially treebanks) makes it difficult to achieve good lexical coverage. Rule-based NLP systems, on the other hand, while not dependent on training data, generally require extensive lexica and/or morphological analyzers as input modules. As it would seem, both statistical and rule-based approaches are resource-sensitive and can run into difficulties with minor languages, or in the face of licensing and financing limitations. What we are addressing in this paper, is the assumption that these problems are especially difficult to solve for rule-based systems, among them our own methodology of choice, Constraint Grammar (CG). At the disambiguational level, early work by Chanod & Tapanainen (1995) showed that even a small set of CG rules can compete with an off-the-shelve statistical tagger. However, in their experiment, both systems had access to an extensive lexicon an a mature morphological analyzer (a finite-state transducer). All published CG parsers to date (Karlsson et al. 1995 and Bick 2000) have included extensive lexico-morphological resources as input to their grammatical rules (as have other rule-based approaches like HPSG and LFG). To the best of our knowledge, no previous research has been done on how to build a CG in the absence of such resources. 2 Related work and methodological focus A great deal of methodological inspiration on the corpus-based creation of lexico- morphological resources can be found in recent research literature, and the different approaches can be clustered into two distinct groups - on the one hand linguistically pre-structured corpus work with an emphasis on paradigm completeness and validation (e.g. Clément, Sagot & Lang 2004 or Forsberg et al. 2006), on the other hand machine learning techniques designed to deduce lexical …

Content truncated to 500 words