Wide-Coverage Semantic Representations from a CCG Parser

Wide-Coverage Semantic Representations from a CCG Parser Johan Bos, Stephen Clark, Mark Steedman School of Informatics, University of Edinburgh � jbos,stevec,steedman � @inf.ed.ac.uk James R. Curran School of Information Technologies, University of Sydney james@it.usyd.edu.au Julia Hockenmaier Institute for Research in Cognitive Science, University of Pennsylvania juliahr@linc.cis.upenn.edu Abstract This paper shows how to construct semantic representations from the derivations produced by a wide-coverage CCG parser. Unlike the de- pendency structures returned by the parser it- self, these can be used directly for semantic in- terpretation. We demonstrate that well-formed semantic representations can be produced for over 97% of the sentences in unseen WSJ text. We believe this is a major step towards wide- coverage semantic interpretation, one of the key objectives of the field of NLP. 1 Introduction The levels of accuracy and robustness recently achieved by statistical parsers (e.g. Collins (1999), Charniak (2000)) have led to their use in a num- ber of NLP applications, such as question-answering (Pasca and Harabagiu, 2001), machine transla- tion (Charniak et al., 2003), sentence simplifica- tion (Carroll et al., 1999), and a linguist’s search engine (Resnik and Elkiss, 2003). Such parsers typically return phrase-structure trees in the style of the Penn Treebank, but without traces and co- indexation. However, the usefulness of this output is limited, since the underlying meaning (as repre- sented in a predicate-argument structure or logical form) is difficult to reconstruct from such skeletal parse trees. In this paper we demonstrate how a wide- coverage statistical parser using Combinatory Cat- egorial Grammar (CCG) can be used to generate se- mantic representations. There are a number of ad- vantages to using CCG for this task. First, CCG provides “surface compositional” analysis of certain syntactic phenomena such as coordination and ex- traction, allowing the logical form to be obtained for such cases in a straightforward way. Second, CCG is a lexicalised grammar, and only uses a small num- ber of semantically transparent combinatory rules to combine CCG categories. Hence providing a com- positional semantics for CCG simply amounts to as- signing semantic representations to the lexical en- tries and interpreting the combinatory rules. And third, there exist highly accurate, efficient and ro- bust CCG parsers which can be used directly for this task (Clark and Curran, 2004b; Hockenmaier, 2003). The existing CCG parsers deliver predicate argu- ment structures, but not semantic representations that can be used for inference. The present paper seeks to extend one of these wide coverage parsers by using it to build logical forms suitable for use in various NLP applications that require semantic in- terpretation. We show how to construct first-order represen- tations from CCG derivations using the λ-calculus, and demonstrate that semantic representations can be produced for over 97% of the sentences in unseen WSJ text. The only other deep parser we are aware of to achieve such levels of robustness for the WSJ is Kaplan et al. (2004). The use of the λ-calculus is integral to our method. However, first-order rep- resentations are simply used as a …

Content truncated to 500 words