‘One’-anaphora and the case for discourse-driven referring expression generation

One-Anaphora and the Case for Discourse-Driven Referring Expression Generation Robert Dale Centre for Language Technology Macquarie University Sydney NSW 2109 Robert.Dale@mq.edu.au Abstract Conventional approaches to the generation of referring expressions place the task within a pipelined ar- chitecture, typically somewhere be- tween text planning and linguis- tic realisation. In this paper, we look at the issues that arise in generating one-anaphoric referring expressions; examination of this task causes us to reflect on the current predominant architectural models for natural language gener- ation, and leads us to suggest an alternative architecture where de- cisions that influence forms of ref- erence happen much earlier in the process of natural language genera- tion. 1 Introduction Referring expression generation is a much- explored task within natural language gen- eration: given an internal symbol that corre- sponds to an entity in some real or imagined world, we need to work out what properties of that entity should be used to describe the en- tity so that our hearer will be able to identify it as the intended referent. Many different al- gorithms have been developed to address this task, which is generally conceived of as map- ping from a symbol—effectively, a referent— to a set of properties—a sense. The com- putation of the appropriate set of properties to use takes account of the other potential referents in the context, selecting properties which rule these distractors out of consider- ation. Conventional approaches to the process of generating referring expressions place the task within a pipelined architecture, where it is assumed that questions of what content should be conveyed in a text are resolved be- fore questions of surface form are considered; this is the well-known strategy vs tac- tics distinction first discussed in the context of natural language generation in the mid- 1970s. However, it is not clear exactly where in the pipeline the process of generating re- ferring expressions should belong. In Reiter and Dale [2002], reflecting current practice in the field, we positioned it in the microplan- ning stage, where microplanning is an inter- mediate stage lying between text planning and surface realisation. Even there, how- ever, we noted that there are interactions be- tween the three microplanning tasks of sen- tence planning, lexical selection and referring expression generation that argue for a more interleaved constraint-based approach to the problem. The principal focus of existing work has been the generation of definite noun phrase refer- ences; relatively little has been written on generating other kinds of referring expres- sions. In particular, there is virtually no work on the generation of one-anaphora. Taking up some ideas first explored in [Dale 1992, 1995], this paper looks at how a considera- tion of where one-anaphora fits into the gen- eration process might cause us to review the kinds of architecture that are required for natural language generation. Section 2 first summarises the conventional approach to referring expression generation, and reviews how this fits into the standard architectural models for natural language generation. Section 3 introduces the phe- …

Content truncated to 500 words