LU konferenceDimensions of Explanatory Value in Computational Models of Natural Language Use.
Prof. Kees van Deemter
Department of Information and Computing Sciences
Utrecht University
The Netherlands
Models of human language use are commonly evaluated in terms of their performance, for instance by means of computational metrics and/or judgements by human judges. In this talk, I will ask what additional criteria there might be in terms of which such models should be evaluated. I will propose a set of criteria that emerge when we view Natural Language Processing as a scientific enterprise whose aim is to explain how speakers and hearers use language; these criteria include generality, parsimony, and support from linguistic or other theories. To illustrate my proposal, I will focus on a research topic that has a long history in NLP and computational modelling, namely Referring Expressions Generation (REG), comparing some recent REG models in terms of the aforementioned criteria. I will conclude by asking whether the same criteria apply to application-oriented NLP as well, and what it might mean for institutional policies if journal editors and conference organisers took my proposal onboard.
K. van Deemter (2023) Dimensions of Explanatory Value in NLP Models. Computational Linguistics 49 (3).