Referring Expression Generation in time-constrained communication

Referring Expression Generation in time-constrained communication Andr´e Costa Mariotti, Ivandr´e Paraboni University of S˜ao Paulo, School of Arts, Sciences and Humanities S˜ao Paulo, Brazil drehmariotti@gmail.com, ivandre@usp.br Abstract In game-like applications and many others, an underlying Natural Language Generation system may have to express urgency or other dynamic aspects of a fast-evolving situation as text, which may be considerably different from text produced under so-called ‘normal’ circumstances (e.g., without time constrains). As a means to shed light on possible differences of this kind, this paper addresses the computational generation of natural language text in time-constrained communication by presenting two experiments that use the attribute selection task of definite descriptions (or Referring Expression Generation - REG) as a working example. In the first experiment, we describe a psycholinguistic study in which human participants are engaged in a time-constrained reference production task. This results in a corpus of time-constrained descriptions to be compared with ‘normal’ descriptions available from an existing (i.e., with no time constraint) REG corpus. In the second experiment, we discuss how a REG algorithm may be customised so as to produce time-constrained descriptions that resemble those produced by human speakers in similar situations. The proposed algorithm is then evaluated against the time-constrained descriptions produced by the human subjects in the first experiment, and it is shown to outperform standard approaches to REG in these conditions. Keywords: Natural Language Generation, Referring Expression Generation, time constraints 1. Introduction In Natural Language Generation (NLG), the subtask of Re- ferring Expression Generation (REG) (Krahmer and van Deemter, 2012) consists of providing linguistic forms to describe (or refer to) discourse objects. REG involves two distinct issues: content determination (or deciding what to say) and surface realisation (or deciding how to say it in a given language). In this work we focus on the former, ad- dressing the issue of attribute selection of definite descrip- tions hereby called REG for simplicity. Of particular interest to the present discussion is the task of generating natural language text from a visual context provided by, e.g., a video game application. One example of application of this kind is illustrated in Figure 11. Figure 1: Example of a referential context In this case, we may assume that an underlying NLG sys- tem would receive as an input the current state of the game, including all relevant objects in the scene and their proper- ties, and would generate context-sensitive messages such as ‘Try to jump as high as possible to avoid X!’ in which X stands for a description of a particular target object r. For 1Image captured from Nintendo’s Super Mario World. instance, assuming r to correspond to the object labelled as R5, the description X may be realised as ‘the little monster under the question mark’, ‘the brown monster at the bot- tom’, ‘the nasty little thing in front of you’, among many others. The choice of contents to be expressed as a definite descrip- tion is of course determined by its communicative goals (Paraboni and van Deemter, 1999). …

Content truncated to 500 words