Abstract

Much work on natural language processing (NLP) in the 1970s and 1980s concentrated on natural language interfaces. Input was analysed to produce a detailed meaning representation that could be interpreted by the underlying application, generally as a database query or as a command. This was tractable because of the constraints imposed by the very limited domains of the applications, but scaling to systems which process more general language proved extremely difficult. Later NLP work developed techniques which were broad-coverage and usable for large scale processing, but too shallow to support the extraction of detailed meaning representations. I’ll explain some of the issues that make associating semantics with arbitrary natural language text a major challenge for computational linguistics and the progress that is now being made towards that goal. I’ll describe some application areas in which broad-coverage semantics is now being used, concentrating on text mining for eScience.