Formalising Pragmatic Information for Natural Language Generation

The main aim of the work presented here is to demonstrate that certain contextual information, traditionally considered to be beyond the scope of semantic representations, can be formalised sufficiently to allow its use in natural language generation (NLG) decisions, in particular syntactic choice. This ‘pragmatic’ information, where the term pragmatic is interpreted as indicating any reference to the context of language use, will also be useful in the interpretation of natural language, and so the representations proposed here will have potential applications throughout natural language processing (NLP). The motivation for the general approach stems from a view of natural language utterances carrying out particular ‘communicative functions’ which must be considered during the production and analysis of any utterance. A ‘split discourse model’ is presented to allow the definition and use of the required functions, the main feature of the model being an explicit separation between a speaker’s own representation of a discourse a hearer’s assumed representation. Certain features of surface form, most importantly intonational features, are interpreted as indicators of the function which a particular utterance is intended to perform. The thesis presents an account of these surface features, both intonational and syntactic, and how they can be selected (and also interpreted) using the split discourse model. The approach will contribute to the ability to generate contextually appropriate language, and also assist in the understanding of real world uses of natural language.