Referring Expression Generation (REG) is a classic task in many Natural Language Generation systems. One aspect of the task is content determination, where an algorithm needs to select which properties of an object to mention in order to identify it for a reader or listener (e.g. "the large red ball next to the lamp"). The development of algorithms in the field has largely been due to a tension between, on the one hand, efficiency (seeking to communicate all and only the information required for identification) and psycholinguistic plausibility (doing this in a way that approaches what human speakers might do in the same situation). Yet there are aspects of these questions that remain under-explored, and which have implications for ongoing debates in both computational REG and psycholinguistics. The present talk seeks to tread the boundary between these two fields, reporting on recent experimental work that seeks to test the predictions of existing algorithms, and on recent computational work that seeks to exploit experimental results with a view to developing algorithms whose output is humanlike, but which can also be viewed as cognitive models that make precise quantitative predictions about speaker behaviour.
Specifically, the talk will address these issues :
1) The extent to which speakers seek to be efficient communicators, compared to their tendency to fall back on "cheap" strategies for content determination that simply rely on property salience and the inherently preferred nature of some properties (such as an object’s colour) over others (such as its size) ;
2) Whether the strategies exhibited by speakers can be captured by current algorithms, and what a more nuanced — and in particular, non-deterministic — algorithm would need to incorporate to capture speaker behaviour ;
3) Whether the selection strategies exhibited by speakers in content selection are "monolithic" — i.e. relatively stable and unchanging — or whether they can be altered as a function of other forces, notably the impact of an interlocutor’s choices in a dialogue context. This also has implications for computational modeling.
