A faster parsing algorithm for Lexicalized Tree-Adjoining Grammars

This paper points out some computational inefficiencies of standard TAG parsing algorithms when applied to LTAGs. We propose a novel algorithm with an asymptotic improvement, from to , where is the input length and are grammar constants that are independent of vocabulary size. Introduction Lexicalized Tree-Adjoining Grammars (LTAGs) were first introduced in (Schabes et al., 1988) as a variant of Tree-Adjoining Grammars (TAGs) (Joshi, 1987). In LTAGs each elementary tree is specialized for some individual lexical item. Following the original proposal, LTAGs have been used in several state-of-the-art, real-world parsers; see for instance (Abeillé & Candito, 2000) and (Doran et al., 2000). Like link grammar (Sleator & Temperley, 1991) and lexicalized formalisms from the statistical parsing literature (Collins, 1997; Charniak, 1997; Alshawi, 1996; Eisner, 1996) LTAGs provide two main recognized advantages over more standard non-lexicalized formalisms: ! subcategorization can be specified separately for each word; and ! each word can restrict the anchors (head words) of its arguments and adjuncts, encoding lexical preferences as well as some effects of semantics and world knowledge. To give a simple example, consider the verb walk, which is usually intransitive but can take an object in some restricted cases. An LTAG can easily specify the acceptability of sentence Mary walks the dog by associating walk with a transitive elementary tree that selects for an indirect object tree anchored at word dog (and some other words within a limited range).