Speaker-Dependent Variation in Content Selection for Referring Expression Generation

Speaker-Dependent Variation in Content Selection for Referring Expression Generation Jette Viethen Centre for Language Technology Macquarie University Sydney, Australia jette.viethen@mq.edu.au Robert Dale Centre for Language Technology Macquarie University Sydney, Australia robert.dale@mq.du.au Abstract In this paper we describe machine learn- ing experiments that aim to characterise the content selection process for distin- guishing descriptions. Our experiments are based on two large corpora of human- produced descriptions of objects in rela- tively small visual scenes; the referring ex- pressions are annotated with their seman- tic content. The visual context of refer- ence is widely considered to be a primary determinant of content in referring expres- sion generation, so we explore whether a model can be trained to predict the col- lection of descriptive attributes that should be used in a given situation. Our exper- iments demonstrate that speaker-specific preferences play a much more important role than existing approaches to referring expression generation acknowledge. 1 Introduction Since at least the late 1980s, referring expression generation (REG) has been a key topic of inter- est in the natural language generation community (see, for example, (Dale, 1989; Dale and Had- dock, 1991; Dale and Reiter, 1995; van der Sluis, 2001; Krahmer and Theune, 2002; Krahmer et al., 2003; Jordan and Walker, 2005; van Deemter, 2006; Gatt and van Deemter, 2006; Kelleher and Kruijff, 2006)); and it has recently served as the focus for the first major evaluation efforts in nat- ural language generation (see, for example, (Belz et al., 2009; Gatt et al., 2009)). This level of atten- tion is due in large part to the consensus view that has arisen as to what is involved in referring ex- pression generation: the task is widely accepted as involving a process of selecting those attributes of an intended referent that distinguish it from other potential distractors in a given context, resulting in what is often referred to as a distinguishing de- scription. Most existing REG algorithms rely on hand- crafted decision procedures whose behaviour is ei- ther entirely deterministic (Dale, 1989; Dale and Haddock, 1991; Gardent, 2002) or can be influ- enced to some degree using parameters such as preference orderings or cost functions over the available properties in order to choose those that should appear in a referring expression (Dale and Reiter, 1995; van der Sluis, 2001; Krahmer and Theune, 2002; Krahmer et al., 2003; van Deemter, 2006; Gatt and van Deemter, 2006; Kelleher and Kruijff, 2006). However, only very limited at- tempts have been made to determine how these parameters should best be instantiated in order to allow an algorithm to mimic human-produced re- ferring expressions. Furthermore, the results of re- cent evaluation exercises (Gupta and Stent, 2005; Viethen and Dale, 2006; Belz and Gatt, 2007; Gatt et al., 2007; Gatt et al., 2008) show that none of these algorithms can be considered an accurate model of human production of referring expres- sions in any of their instantiations. In this paper, we take a speaker-oriented per- spective on REG that is aimed in part …

Content truncated to 500 words