Experiments with Corpus-based LFG Specialization

Experiments with Corpus-based LFG Specialization Nicola Cancedda and Christer Samuelsson Xerox Research Centre Europe, 6, chemin de Maupertuis 38240 Meylan, France {cancedda.lsamuelsson}@xrce.xerox.com Abstract Sophisticated grammar formalisms, such as LFG, al- low concisely capturing complex linguistic phenom- ena. The powerful operators provided by such for- malisms can however introduce spurious ambigu- ity, making parsing inefficient. A simple form of corpus-based grammar pruning is evaluated experi- mentally on two wide-coverage grammars, one En- giish and one French. Speedups of up to a factor 6 were obtained, at a cost in grammatical coverage of about 13%. A two-stage architecture allows achiev- ing significant speedups without introducing addi- tional parse failures. 1 Introduction Expressive grammar formalisms allow grammar de- velopers to capture complex linguistic generaliza- tions concisely and elegantly, thus greatly facilitat- ing grammar development and maintenance. (Car- rol, 1994) found that the empirical performance when parsing with unification-based grammars is nowhere near the theoretical worst-case complexity. Nonetheless, directly parsing with such grammars, in the form they were developed, can be very ineffi- cient. For this reason, grammars are typically com- piled into representations that allow faster parsing. This does however not solve the potential problem of the grammars overgenerating considerably, thus allowing large amounts of spurious ambiguity. In- deed, a current trend in high-coverage parsing, es- pecially when employing a statistical model of lan- guage, see, e.g., (Collins 97), is to allow the grammar to massively overgenerate and instead disambiguate by statistical means during or after parsing. If the benefits resulting from more concise grammatical de- scriptions are to outweigh the costs of spurious am- biguity, the latter must be brought down. In such a situation, corpus-based compilation techniques can drastically improve parsing perfor- mance without burdening the grammar developer. The initial, and much seminal work in this area was been carried out by Rayner and coworkers, see (Rayner 1988), (Samuelsson and Rayner 91) and (Rayner and Carter 1996). In the current article, we apply similar ideas to Lexical Functional Gram- mar (LFG) in the incarnation of the Xerox Linguis- tic Environment (XLE). The goal is to investigate to what extent corpus-based compilation techniques can reduce overgeneration and spurious ambiguity, and increase parsing efficiency, without jeopardiz- ing coverage. The rest of the article is organized as follows: Section 2 presents the relevant aspects of the LFG formalism and the pruning strategy em- ployed, Section 3 describes the experimental setup, Section 4 reports the experimental results and Sec- tion 5 relates this to other work. 2 LFG and Grammar Pruning The LFG formalism (Kaplan and Bresnan, 1982) al- lows the right-hand sides (RHS) of grammar rules to consist of a regular expression over grammar sym- bols. This makes it more appropriate to refer to the grammar rules as rule schemata, since each RHS can potentially be expanded into a (possibly infinite) number of distinct sequences of grammar symbols, each corresponding to a traditional phrase-structure rule. As can easily be imagined, the use of regular- expression operators such as Kleene-star and com- plementation may introduce …

Content truncated to 500 words