It is a truism of NLG that good knowledge of the reader can improve the quality of generated texts, and many NLG systems have been developed that exploit detailed user models when generating texts. Unfortunately, it is very difficult in practice to obtain detailed information about users. In this paper we describe our experiences in acquiring and using limited user models for NLG in four different systems, each of which took a different approach to this issue. One general conclusion is that it is useful if imperfect user models are understandable to users or domain experts, and indeed perhaps can be directly edited by them; this agrees with recent thinking about user models in other applications such as intelligent tutoring systems (Kay, 2001).
|Number of pages||8|
|Publication status||Published - 2003|
|Event||9th European Workshop on Natural Language Generation, ENLG@EACL 2003 - Budapest, Hungary|
Duration: 13 Apr 2003 → 14 Apr 2003
|Conference||9th European Workshop on Natural Language Generation, ENLG@EACL 2003|
|Period||13/04/03 → 14/04/03|