Content Planning for Neural Story Generation with Aristotelian Rescoring

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4319–4338, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4319 Content Planning for Neural Story Generation with Aristotelian Rescoring Seraphina Goldfarb-Tarrant∗†, Tuhin Chakrabarty∗‡, Ralph Weischedel∗, Nanyun Peng∗§ ∗ Information Sciences Institute, University of Southern California † University of Edinburgh, ‡ Columbia University, § University of California, Los Angeles s.tarrant@ed.ac.uk, tuhin.chakr@cs.columbia.edu, weisched@isi.edu, violetpeng@cs.ucla.edu Abstract Long-form narrative text generated from large language models manages a fluent imperson- ation of human writing, but only at the local sentence level, and lacks structure or global co- hesion. We posit that many of the problems of story generation can be addressed via high- quality content planning, and present a system that focuses on how to learn good plot struc- tures to guide story generation. We utilize a plot-generation language model along with an ensemble of rescoring models that each im- plement an aspect of good story-writing as detailed in Aristotle’s Poetics. We find that stories written with our more principled plot- structure are both more relevant to a given prompt and higher quality than baselines that do not content plan, or that plan in an unprin- cipled way.1 1 Introduction Despite many recent advances in Natural Language Generation, successful creative narrative composi- tion remains elusive. Current neural approaches are plagued by difficulty in mastering structure, will veer between topics, and lack long-range cohesion. They successfully imitate the fluency and style of human writing, but on closer inspection sentences do not fit together to form a whole, and the reader is left with the impression that the generation has no content (See et al., 2019). This lack of structure also degrades the relevance of generations condi- tioned on a prompt or other source text - a strong language model will repeat key phrases from a given prompt but will not remain on topic. These issues are illustrated in the Naive Generated Story in Table 1, where many of the sentences individually are fine, but do not fit together as one story, and do not all relate to the prompt. 1Code at https://github.com/PlusLabNLP/ story-gen-BART We hypothesise that this problem can be ad- dressed with a focus on deeper latent narrative structures. In Aristotle’s Poetics, one of the most enduring treatises on the craft of writing good sto- ries, the philosopher lays out the elements of story in order of importance. They are: 1. event choice and arrangement 2. character 3. relevant content2 4. diction An amateur masters skills later in the list, but mas- tery of event choice and event arrangement is what distinguishes a good writer (Aristotle). Next is character, then relevance, and only finally do style and diction matter. This philosophical framework fits remarkably well into the traditional Natural Language Gener- ation Pipeline approach that emphasizes Content Planning (Reiter and Dale, 1997). The pipeline di- vides generation into three steps: Content Planning, Microplanning and Surface Realization, where at each step input is modified and refined, getting closer to the final textual output. …

Content truncated to 500 words