The Importance of Supertagging for Wide-Coverage CCG Parsing

The Importance of Supertagging for Wide-Coverage CCG Parsing Stephen Clark School of Informatics University of Edinburgh 2 Buccleuch Place, Edinburgh, UK stephen.clark@ed.ac.uk James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia james@it.usyd.edu.au Abstract This paper describes the role of supertagging in a wide-coverage CCG parser which uses a log-linear model to select an analysis. The supertagger reduces the derivation space over which model estimation is performed, reducing the space required for discriminative training. It also dramatically increases the speed of the parser. We show that large increases in speed can be obtained by tightly integrating the su- pertagger with the CCG grammar and parser. This is the first work we are aware of to success- fully integrate a supertagger with a full parser which uses an automatically extracted grammar. We also further reduce the derivation space us- ing constraints on category combination. The result is an accurate wide-coverage CCG parser which is an order of magnitude faster than com- parable systems for other linguistically moti- vated formalisms. 1 Introduction Lexicalised grammar formalisms such as Lexical- ized Tree Adjoining Grammar (LTAG) and Com- binatory Categorial Grammar (CCG) assign one or more syntactic structures to each word in a sentence which are then manipulated by the parser. Supertag- ging was introduced for LTAG as a way of increasing parsing efficiency by reducing the number of struc- tures assigned to each word (Bangalore and Joshi, 1999). Supertagging has more recently been applied to CCG (Clark, 2002; Curran and Clark, 2003). Supertagging accuracy is relatively high for man- ually constructed LTAGs (Bangalore and Joshi, 1999). However, for LTAGs extracted automati- cally from the Penn Treebank, performance is much lower (Chen et al., 1999; Chen et al., 2002). In fact, performance for such grammars is below that needed for successful integration into a full parser (Sarkar et al., 2000). In this paper we demonstrate that CCG supertagging accuracy is not only suffi- cient for accurate and robust parsing using an auto- matically extracted grammar, but also offers several practical advantages. Our wide-coverage CCG parser uses a log-linear model to select an analysis. The model paramaters are estimated using a discriminative method, that is, one which requires all incorrect parses for a sen- tence as well as the correct parse. Since an auto- matically extracted CCG grammar can produce an extremely large number of parses, the use of a su- pertagger is crucial in limiting the total number of parses for the training data to a computationally manageable number. The supertagger is also crucial for increasing the speed of the parser. We show that spectacular in- creases in speed can be obtained, without affecting accuracy or coverage, by tightly integrating the su- pertagger with the CCG grammar and parser. To achieve maximum speed, the supertagger initially assigns only a small number of CCG categories to each word, and the parser only requests more cate- gories from the supertagger if it cannot provide an analysis. We also demonstrate how extra constraints on the …

Content truncated to 500 words