Abstract

For a large portion of its now substantial life, natural language processing (NLP) has developed under two restricting assumptions: that research should focus on subtasks until these are mature enough to be integrated, and that research should concentrate on easy texts before addressing more complex ones. These restrictions are based on sound engineering principles and they have allowed NLP as a research field to survive in spite of the difficulty of the overall task. The current level of maturity of some of these subtasks, and the glass ceilings that many of them appear to have reached suggest that it may be time to venture a little beyond. Research on the generation of literary texts provides the opportunity to do so. It forces the consideration of several of these subtasks in cooperation, and it requires the analysis of complex phenomena, such as narrative, figurative language, or linguistic creativity that are usually avoided in NLP . This talk will address some of these open problems and outline their potential and their difficulties. It will show how engineering considerations suggest that language analysis and language generation should be considered in tandem to address these problems, and how some currently accepted architectures may need to be revised to accommodate some of these phenomena. To illustrate these points, the current scope and architecture of the WASP poetry generation system will be described and discussed. Over the last ten years, the WASP system has been substantially overhauled both to expand the range of phenomena it considers and to integrate as many as possible of the available NLP technologies. Although this is work in progress, it already shows some interesting insights concerning system architecture and knowledge representation. These insights will be described, and open lines of future work will be discussed.