A Finite-State Parser with Dependency Structure Output

Dependency parsers and nite-state parsers are both capable of rapid and robust parsing of natural language. Dependency parsers produce richer output structures, while nite-state parsers can be more eecient. We show how a nite-state parser can be used to produce dependency structures for most phrase types, with an O(n 2) complexity in the number of words. The parser allows syntactically ambiguous structures to be packed into a single representation. The parser has been used as a component of a natural-language information retrieval system, operating in English and Japanese. 1 Dependency parsing and nite-state parsing Recent work in parsing natural language has shown a shift away from the heavyweight theories of syntax and semantics, towards grammar formalisms which can be parsed eeciently and which make it easy to write robust grammars. Many of the lightweight techniques use either nite-state parsing, or some variant of dependency grammars. Finite-state parsing ooers the advantages of computational eeciency, and the ability to easily integrate a number of levels of processing, from the phonological (Bird and Ellison, 1994) through to the syntactic (Roche, 1997; Roche and Schabes, 1997). Dependency grammars have a long history in linguistics and NLP, from the linguistically inclined work of Mel' cuk (1988), to computation implementations such as the dependency parser of JJ arvinen and Tapanainen (1997). A related kind of grammar is the Link grammar of Sleator and Temperley (1991). The processing complexity of dependency grammars can be as low as O(n 3) (Eisner, 1996), where n is the number of words in the input. However, as Neuhaus and Brr oker (1997) show, if certain sorts of linguistic construct are allowed, the processing complexity becomes exponential, up to being NP-complete. The speciic constructs that cause this are non-projective ones, in which links between words must cross over, i.e. if there is a sequence of words W i :::W j :::W k in which W i and W k are linked, then W j has a link to a word preceding W i or following W k. Neuhaus and Brr oker show that such constructs can be found in English, in the form of topicalisation, and in German. Similar problems arise when the parser is allowed to tolerate ungrammatical input. For example, if W i and W k were spurious words in the input (say the