An Evaluation of Lexicalization in Parsing

In this paper , we evaluate a two-pass parsing strategy proposed for the so-called ' lexicalized' grammar. In ' lexicalized' grammars (Schabes, Abeill$ and Joshi, 1988), each elementary structure is systematical ly associated with a lexical i tem called anchor. These structures specify extended domains of locality (as compared to CFGs) over which constraints can be stated. The 'g rammar ' consists of a lexicon where each lexical i tem is associated with a finite number of structures for which that i tem is the anchor. There are no separate grammar rules. There are, of course, ~rules' which tell us how these structures are combined. A general two-pass parsing strategy for ' lexicalized' grammars follows naturally. In the first stage, the parser selects a set of elementary structures associated with the lexical i tems in the input sentence, and in the second stage the sentence is parsed with respect to this set. We evaluate this s trategy with respect to two characteristics. First , the amount of filtering on the entire grammar is evaluated: once the first pass is performed, the parser uses only a subset of the grammar. Second, we evaluate the use of non-local information: the structures selected during the first pass encode the morphological value (and therefore the position in the string) of their anchor; this enables the parser to use non-local information to guide its search. We take Lexicalized Tree Adjoining Grammars as an instance of lexicallzed grammar. We i l lustrate the organization of the grammar. Then we show how a general Earley-type TAG parser (Schabes and Joshi, 1988) can take advantage of lexicalization. Empirical da ta show that the filtering of the grammar and the non-local information provided by the two-pass strategy improve the performance of the parser. 1 L E X I C A L I Z E D G R A M M A R S Most cu r ren t l inguis t ic theor ies give lexical accounts of severa l p h e n o m e n a t h a t used to be cons idered pure ly syn tac t i c . T h e in fo rma t ion p u t in the lexicon is t he reby increased in b o t h a m o u n t and complex i ty : see, for example , lexical rules in L F G ( K a p l a n and Bresnan , 1983), G P S G ( G a z d a r , Kle in , P u l l u m and Sag, 1985), H P S G (Po l l a rd and Sag, 1987), C o m b i n a t o r y Ca tego r i a l G r a m m a r s ( S t e e d m a n 1985, 1988), K a r t t u n e n ' s vers ion of Ca t ego r i a l G r a m m a r ( K a r t t u n e n 1986, 1988), some vers ions of GB theo ry ( C h o m s k y 1981), and L e x i c o n G r a m m a r s (Gross 1984). We say t h a t a g r a m m a r is ' l ex ica l ized ' if i t consis ts of.. 1 • a f ini te set of s t r u c t u r e s each as soc ia t ed wi th a lexical i tem; each lexical i t e m will be cal led the anchor of the co r r e spond ing s t ruc tu re ; t he s t ruc tu re s define the d o m a i n of loca l i ty over which cons t r a in t s are specified; cons t ra in t s are local wi th r e spec t to the i r anchor; • an o p e r a t i o n or ope ra t i ons for compos ing the s t ruc tu res . Not ice t h a t Ca t ego r i a l G r a m m a r s (as used for example by Ades and S t e e d m a n , 1982 and S t e e d m a n , 1985 and 1988) are ' l ex ica l ized ' accord ing to our def ini t ion since each bas ic ca t egory has a lexical i t e m assoc ia ted wi th it. A genera l two-s tep pa r s ing s t r a t e g y for ' l ex ica l i zed ' g r a m m a r s follows na tu ra l ly . In t he first s tage , the pa r se r selects a set of e l e m e n t a r y s t ruc tu re s assoc ia ted wi th the lexical i t ems in the i npu t sentence , and in the second s tage the sentence is pa r sed wi th r e spec t to th is set . T h e s t r a t e g y is i n d e p e n d e n t of the na tu r e of the e l e m e n t a r y s t ruc tu re s in the unde r ly ing g r a m m a r . In pr incip le , any pa r s ing a l g o r i t h m can be used in the second s tage. 1 By qexicalization' we mean that in each structure there is a lexical item that is realized. We do not mean simply adding feature structures (such as head) and unification equations to the rules of the formalism.