CCG Parsing

Introduction

The ZPar CCG parser is essentially the ZPar generic version of the phrase-structure parser with special processing of lexical categories and combinatorial rules (including unary rules). The source code are located at zpar/src/common/conparser/GENERIC_CCGPARSER_IMPL. The implementation macro is located at zpar/Makefile.ccg.

To compile the parser, type make generic.ccgparser. The target binary zpar/dist/generic.ccgparser/train takes three additional arguments compared to the phrase-structure parsers.

• -b binary rules (example)
• -u unary rules (example)
• -c lexical category files (example)

Usage of the target binary zpar/dist/generic.ccgparser/conparser is similar to the English phrase-structure parser, except that the input file contains lexical categories in addition to POS (an example).

The output files are binarized derivations, in the same format as conparser. Here are an example training file and an example reference file.

The CCG parser is tuned in a similar way to the conparser, where a certain number of training iterations are performed and the best iteration is decided using a set of development data. Here is an example script for tuning experiments. Note that the C&C parser is used to convert derivations to CCG dependencies, which are the normal objectives of evaluations.

Reference

• Yue Zhang and Stephen Clark. 2011. Shift-Reduce CCG Parsing. In Proc. of ACL, pages 683-692.