Statistical and logical reasoning in disambiguation
Stephen Pulman · Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences · 2000
While recent statistical methods for disambiguation in natural language processing have been very successful, earlier arguments still hold for the position that fully successful disambiguation requires reasoning to new conclusions from old facts. We explore ways of complementing statistical approaches with the use of ‘domain theories’: collections of facts and axioms that characterize the typical structure of some task domain. In particular, we hypothesize that disambiguation decisions can supply tacit information about such theories, and that the theories can, in part, be automatically induced from such data. We describe a pilot experiment in which a partial domain theory for the domain of air–travel information was induced from a corpus of disambiguated example sentences, the resulting theory then being used successfully in disambiguating other sentences from the same domain.