Computational modelling of contextual coreference: implications for Swedish text-to-speech
Merle A. Horne, Marcus Filipsson, Mats Ljungqvist, Anders Lindström · 1994
The issue of describing and identifying coreference relations between lexical (content) words is discussed. In textto -speech applications, it is important to be able to recognize these relations since contextually coreferent (given) words are associated with tone accent patterns that differ from those on new information. An algorithm for referent tracking in a restricted domain is also described which allows one to preprocess a text and automatically tag words as either contextually `New' or `Given'. The algorithm presupposes computational modelling of lexical semantic identity of sense relations as well as information on inflexional/derivational morphology and compounding. This information is available in a lemmatized lexicon of Swedish. Referent identity is defined on head-word representations derived from the text input on the basis of the inflexional expansion rules contained in the lexicon. Information on the New/Given status of words can subsequently be used in the F0-generating...