Narrative Composition: Achieving the Perceived Linearity of Narrative

The last few years have seen an increased interest in narrative within the field of Natural Language Generation (Reiter et al., 2008; Elson and McKeown, 2010; Siddharthan et al., 2012; Lester, 2012). Narrative is generally acknowledged as a fundamental mode of presenting and communicating information between humans, with different manifestations across media but with a very significant presence in textual form. Yet efforts in Natural Language Generation research have generally side stepped the issue. Aside from the pioneering work of (Callaway, 2002) and an early attempt to bridge the gap between narratology and natural language generation (Lönneker, 2005), the field had mostly avoided narrative until recent times. Two possible arguments may be considered as an explanation of this: one based on the need to restrict initial work within a field to the simpler challenges before tackling the difficult ones, and another based on an assumption that the peculiarities of narrative have already been covered by existing work. Both arguments can be shown to be inappropriate. With respect to the first argument, the field of natural language generation has for many years operated under the tacit assumption that state of the art technology can only aspire to generating texts within a limited range of domains and genres. These have over the years been defined in different ways, but in spite of changes, literary texts have usually been considered to be outside the range of possible candidates. From an engineering point of view, this kind of restriction made sense when the field was starting, for two important reasons. One, the technological solutions available at the time for the various tasks involved in natural language generation were in their infancy, and the linguistic complexity of literary text might have been beyond their scope. Two, natural language generation arose from a desire to extend the studies that had been carried out for computational analysis of language to the task of generation, and what was known about language from a computational point of view concerned simple texts. Most of the studies on language and computation had applied similar simplifying assumptions. However, such restricting assumptions are no longer necessary and may be inhibiting progress. In terms of technology, the field has matured significantly over the intervening years. The current state of the art provides a wide range of solutions that may be well suited to address some of the more complex phenomena involved in literary text. Additional objections may be made on the grounds that we do not know enough about these phenomena. Such objections, however valid they might have been originally, are no longer valid either. Many of the phenomena that were considered beyond computational treatment (metaphor, emotion, temporal reasoning, dialogue...) have been the subject of serious and sustained study over the same time period. Many approaches to their computational modelling and treatment have become available. More to the point, the last few years have seen a rise of interest on literary text within the natural language processing community. This is evidenced by the number of workshops addressing topics related to literature: Workshop on Computational Approaches to Linguistic Creativity at NAACL HLT 2009 and 2010, Computational Linguistics for Literature Workshop at NAACL HLT 2012 and 2013, Computational Models of Narrative events held as AAAI Fall symposium in 2010, as LREC workshop in 2012, and as satellite workshop of CogSci 2013, just to name a few. With respect to the second argument, the recent reappearance of narrative as a research topic for NLG should be enough to dispel the notion that all its problems have already been solved. Narrative has many peculiarities that set it apart from other kinds of text, and the body of work addressing narrative as a research topic within NLG has