ML-Tuned Constraint Grammars

ML-Tuned Constraint Grammars Eckhard Bick Institute of Language and Communication University of Southern Denmark, Odense eckhard.bick@mail.dk Abstract In this paper we present a new method for machine learning-based optimization of linguist-written Constraint Grammars. The effect of rule ordering/sorting, grammar- sectioning and systematic rule changes is discussed and quantitatively evaluated. The F-score improvement was 0.41 percentage points for a mature (Danish) tagging grammar, and 1.36 percentage points for a half-size grammar, translating into a 7-15% error reduction relative to the performance of the untuned grammars. 1 Introduction Constraint Grammar (CG) is a rule-based paradigm for Natural Language Parsing (NLP), first introduced by Karlsson et al. (1995). Part- of-speech tagging and syntactic parses are achieved by adding, removing, selecting or substituting form and function tags on tokens in running text. Rules express linguistic contextual constraints and are written by hand and applied sequentially and iteratively, ordered in batches of increasing heuristicity and incrementally reducing ambiguity from morphologically analyzed input by removing (or changing) readings from so-called readings cohorts (consisting of all possible readings for a given token), - optimally until only one (correct) reading remains for each token. The method draws robustness from the fact that it is reductionist rather than generative - even unforeseen or erroneous input can be parsed by letting the last reading survive even if there are rules that would have removed it in a different context. Typical CG rules consist of an operator (e.g. REMOVE, SELECT), a target and one or more contextual constraints that may be linked to each other: (a) REMOVE VFIN (-1C ART OR DET) ; (b) SELECT VFIN (-1 PERS/NOM) (NOT *1 VFIN) Rule (a), for instance, removes a target finite verb reading (VFIN) if there is an unambiguous (C) article or determiner 1 position to the left (-), while rule (b) selects a finite verb reading, if there is a personal pronoun in the nominative immediately to the left, and no (NOT) other finite verb is found anywhere to the right (*1). Mature Constraint Grammars can achieve very high accuracy, but contain thousands of rules and are expensive to build from scratch, traditionally requiring extensive lexica and years of expert labor. Since grammars are not data- driven in the statistical sense of the word, domain adaptation, for instance for speech (Bick 2011) or historical texts (Bick 2005), is traditionally achieved by extending an existing general grammar for the language in question, and by using specialized lexica or two-level text normalization. However, due to its innate complexity, the general underlying grammar as a whole has properties that do not easily lend themselves to manual modification. Changes and extensions will usually be made at the level of individual rules, not rule interactions or rule regrouping. Thus, with thousands of interacting rules, it is difficult for a human grammarian to exactly predict the effect of rule placement, i.e. if a rule is run earlier or later in the sequence. In particular, rules with so-called C-conditions (asking for unambiguous context), may profit from another, earlier rule …

Content truncated to 500 words