Abstract

In this talk I will describe a state-of-the-art parser of natural language which uses a grammatical formalism from the theoretical linguistics literature: Combinatory Categorial Grammar (CCG). CCG was designed to capture unbounded dependencies in text, such as those arising from coordination and extraction phenomena.

I will describe a recent evaluation which measures how well the CCG parser – and other off-the-shelf parsers – are able to recover such dependencies. I will also show that the use of CCG leads to a surprisingly robust and efficient parser. The parser is able to analyse sentences of naturally occurring text – from newspapers and biomedical research papers, for example – at speeds of tens of sentences per second. This allows practical generation of sophisticated linguistic representations on a very large scale.