Learning to Tell Tales: A Data-driven Approach to Story Generation

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 217–225, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Learning to Tell Tales: A Data-driven Approach to Story Generation Neil McIntyre and Mirella Lapata School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh, EH8 9AB, UK n.d.mcintyre@sms.ed.ac.uk, mlap@inf.ed.ac.uk Abstract Computational story telling has sparked great interest in artificial intelligence, partly because of its relevance to educa- tional and gaming applications. Tradition- ally, story generators rely on a large repos- itory of background knowledge contain- ing information about the story plot and its characters. This information is detailed and usually hand crafted. In this paper we propose a data-driven approach for gen- erating short children’s stories that does not require extensive manual involvement. We create an end-to-end system that real- izes the various components of the gen- eration pipeline stochastically. Our system follows a generate-and-and-rank approach where the space of multiple candidate sto- ries is pruned by considering whether they are plausible, interesting, and coherent. 1 Introduction Recent years have witnessed increased interest in the use of interactive language technology in ed- ucational and entertainment applications. Compu- tational story telling could play a key role in these applications by effectively engaging learners and assisting them in creating a story. It could also al- low teachers to generate stories on demand that suit their classes’ needs. And enhance the enter- tainment value of role-playing games1. The major- ity of these games come with a set of pre-specified plots that the players must act out. Ideally, the plot should adapt dynamically in response to the play- ers’ actions. Computational story telling has a longstanding tradition in the field of artificial intelligence. Early work has been largely inspired by Propp’s (1968) 1A role-playing game (RPG) is a game in which the par- ticipants assume the roles of fictional characters and act out an adventure. typology of narrative structure. Propp identified in Russian fairy tales a small number of recurring units (e.g., the hero is defeated, the villain causes harm) and rules that could be used to describe their relation (e.g., the hero is pursued and the rescued). Story grammars (Thorndyke, 1977) were initially used to capture Propp’s high-level plot elements and character interactions. A large body of more recent work views story generation as a form of agent-based planning (Theune et al., 2003; Fass, 2002; Oinonen et al., 2006). The agents act as characters with a list of goals. They form plans of action and try to fulfill them. Interesting stories emerge as agents’ plans interact and cause failures and possible replanning. Perhaps the biggest challenge faced by compu- tational story generators is the amount of world knowledge required to create compelling stories. A hypothetical system must have information about the characters involved, how they inter- act, what their goals are, and how they influence their environment. Furthermore, all this informa- tion must be complete and error-free if it is to be used as input …

Content truncated to 500 words