Learning Phrasal Categories

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 301–307, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Learning Phrasal Categories William P. Headden III, Eugene Charniak and Mark Johnson Brown Laboratory for Linguistic Information Processing (BLLIP) Brown University Providence, RI 02912 {headdenw|ec|mj}@cs.brown.edu Abstract In this work we learn clusters of contex- tual annotations for non-terminals in the Penn Treebank. Perhaps the best way to think about this problem is to contrast our work with that of Klein and Man- ning (2003). That research used tree- transformations to create various gram- mars with different contextual annotations on the non-terminals. These grammars were then used in conjunction with a CKY parser. The authors explored the space of different annotation combinations by hand. Here we try to automate the process — to learn the “right” combination auto- matically. Our results are not quite as good as those carefully created by hand, but they are close (84.8 vs 85.7). 1 Introduction and Previous Research It is by now commonplace knowledge that accu- rate syntactic parsing is not possible given only a context-free grammar with standard Penn Tree- bank (Marcus et al., 1993) labels (e.g., S, NP, etc.) (Charniak, 1996). Instead researchers condition parsing decisions on many other fea- tures, such as parent phrase-marker, and, fa- mously, the lexical-head of the phrase (Mager- man, 1995; Collins, 1996; Collins, 1997; Johnson, 1998; Charniak, 2000; Henderson, 2003; Klein and Manning, 2003; Matsuzaki et al., 2005) (and others). One particularly perspicuous way to view the use of extra conditioning information is that of tree-transformation (Johnson, 1998; Klein and Manning, 2003). Rather than imagining the parser roaming around the tree for picking up the infor- mation it needs, we rather relabel the nodes to di- rectly encode this information. Thus rather than have the parser “look” to find out that, say, the parent of some NP is an S, we simply relabel the NP as an NP[S]. This viewpoint is even more compelling if one does not intend to smooth the probabilities. For example, consider p(NP → PRN | NP[S]) If we have no intention of backing off this probabil- ity to p(NP → PRN | NP) we can treat NP[S] as an uninterpreted phrasal category and run all of the standard PCFG algorithms without change. The result is a vastly simplified parser. This is ex- actly what is done by Klein and Manning (2003). Thus the “phrasal categories” of our title refer to these new, hybrid categories, such as NP[S]. We hope to learn which of these categories work best given that they cannot be made too specific because that would create sparse data problems. The Klein and Manning (2003) parser is an un- lexicalized PCFG with various carefully selected context annotations. Their model uses some par- ent annotations, and marks nodes which initiate or in certain cases conclude unary productions. They also propose linguistically motivated annotations for several tags, including V P, IN, CC,NP and S. This results in a …

Content truncated to 500 words