Abstract

In this talk we will describe two approaches to improving the efficiency of a wide-coverage CCG parser. The C&C parser (Clark and Curran, 2007) is already surprisingly efficient, parsing at around 30 sentences per second on standard hardware. This is surprising because the parser does not do any pruning at the parsing stage, but builds a complete packed chart. The efficiency comes from the use of a linear-time supertagger, which greatly reduces the search space, and highly optimised C++.

Despite the use of the supertagger, there is still a huge amount of ambiguity left in the chart. The first approach to improving speed will be to perform some pruning on the chart. We investigate standard beam search for chart-parsing—removing low-scoring items—as well as a novel, more aggressive technique which removes complete cells from the chart. Both techniques result in significant speed-ups.

The second approach is a novel technique which involves self-training the supertagger on large amounts of parser output. The speed of the parser is directly related to the number of supertags (CCG lexical categories) supplied by the supertagger for each word on average. The insight behind this approach is to recognise that the supertagger can easily be trained to predict which supertags the parsing model will eventually choose, resulting in a supertagger model which is much more tightly integrated with the parser.