The Design of an Experiment in Anaphora Resolution for Referring Expressions Generation

International Conference RANLP 2009 - Borovets, Bulgaria, pages 225–229 The Design of an Experiment in Anaphora Resolution for Referring Expressions Generation Diego Jesus de Lucena University of São Paulo (USP/EACH) Av. Arlindo Bettio, 1000 São Paulo, Brazil diego.si@usp.br Ivandré Paraboni University of São Paulo (USP/EACH) Av. Arlindo Bettio, 1000 São Paulo, Brazil ivandre@usp.br Abstract We present a pilot experiment to measure the effects of redundancy in the resolution of definite descriptions as performed by a small number of human readers. Although originally intended to provide evidence of how much redundancy should ideally be included in generated anaphoric descriptions, preliminary findings reveal a number of little explored issues that are relevant to both referring expressions generation and interpretation. Keywords Referring expressions generation, Anaphora resolution. 1. Introduction Human speakers routinely make use of redundant information when referring to world or discourse objects via definite descriptions, and they often do so even when the sole purpose of referring is the identification of the target object. By contrast, Natural Language Generation (NLG) systems usually implement referring expressions generation (REG) algorithms that are far less prepared to include redundant information in their output. One possible reason for this difference between real language use and NLG output is the fact that generating redundancy without a proper reason comes with a price, namely, false logic implicatures in the sense defined by H. P. Grice [1]. For instance, in a context containing only one object of type ‘door’, a redundant (from the point of view of identification) reference to the colour attribute of the referent as in “please open the red door” may cause the hearer to consider whether there is any special reason for mentioning such ‘unnecessary’ information at all. To avoid this sort of mishap, most REG algorithms to date (including one of the most influential works in the field, the Incremental algorithm in [2]) attempt to avoid the inclusion of any information not strictly necessary for the identification of the intended referent. Referring expressions produced in this way are suitably brief, but they may look unnaturally so. In extreme cases, certain instances of short descriptions may actually make the identification of the intended referent a daunting task. One such example is the case of deictic references in structurally-complex (e.g., spatial) domains. Deictic referents may be unidentifiable unless a certain amount of redundant information is added [3]. For example, a distinguishing description such as “the girl in white shoes” is not of much help if, say, the referred person is part of a large crowd. Redundancy in this case (e.g., “the girl in white shoes, next to the elevator”) may facilitate the resolution of these expression (here understood as the task of interpreting the referring expression and identifying the intend referent.) The implication of this for REG algorithms is that redundancy should be somehow taken into account at least when generating instances of space deixis, and this is precisely the kind of insight needed to design NLG systems that describe objects in physical contexts. For other …

Content truncated to 500 words