Character Networks for Narrative Generation

In this position paper, the author proposes the use of social networks of characters as an AI narrative generation mechanism. The first part of the paper offers examples of recent research by literary critics on the relationship between character networks and narrative structure. The second part of the paper offers a simple example of story generation based on a structural balance network model. From Narratives to Networks Finding realistic but tractable story generation mechanisms is an ongoing challenge for the narrative intelligence community. Most story generators rely on either (1) corpora of pre-existing stories (e.g., MEXICA), or (2) story grammars (e.g., TALESPIN). Efforts have been made to broaden generative mechanisms to include games as well as crowdsourcing. In a recent paper, Pablo Gervas argues for the value of chess as a narrative generation mechanism: Chess provides a finite set of characters (pieces), a schematical representation of space (the board), and time (progressive turns), and a very restricted set of possible actions. Yet it also allows very elementary interpretations of game situations in terms of human concepts such as danger, threat, conflict, death, survival, victory or defeat, which can be seen as interesting building blocks for story construction. There is a related body of work that makes use of sports games statistics to generate simple narratives akin to newspaper articles (Allen et al, 2010; Laureau et al, 2011; Copyright © 2012, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. 1 Perez y Perez, R., Sharples, M. MEXICA. Journal of Experimental and Theoretical Artificial Intelligence, 13. 2006. pp. 121-139 2 Meehan, J. Tale-spin. In Proceedings of IJCAI 1977. pp. 91-98. 3 See Reidl, Mark et al. “Automatically Learning to Tell Stories about Social Situations from the Crowd” CMN Proceedings. LREC 2012. 4 Gervas, Pablo. “From Fleece of Fact to Narrative Yarns.” Computational Models of Narrative Proceedings, LREC 2012. pp. 123. Bouayad-Agha et al, 2011). Basic simulations have been used similarly. Computational social scientist Robert Axtell argues that each run of Sugarscape (an agent-based model of wealth and trade) constitutes an historical narrative (or “proto-history”) describing the social and cultural development of a population of artificial agents. In the position paper that follows, I will propose that social networks may provide another, complementary mechanism for narrative generation. The relationship between social networks and narrative structure has been an active area of research for literary critics in the past several years. Efforts have been directed particularly at the extraction of character interaction networks from literary works. The guiding principle behind literary network analysis is that narratives are not merely depictions of individual experience in language but are also artificial societies whose imaginary social forms can be quantified and analyzed. What such analyses reveal is that narrative structure (e.g., plot, genre, and characterization) is intimately related to network structure. Included below are examples drawn from my own research on networks of characters in 19 Century British fiction. Figure 1 shows the social network diagrams for Charles Dickens’ The Pickwick Papers, George Eliot’s Middlemarch, and Henry James The Ambassadors. 5 Axtell, Robert and Epstein, Joshua. Growing Artificial Societies. Washington: Brookings, 1996. 6 Dames N., Elson D., McKeown K., “Extracting Social Networks from Literary Fiction.” Assoc. Computational Linguistics Proceedings, 2010. 7 Sack, Graham. Character Networks in the 19 Century British Novel. Columbia University, MA Thesis, 2011. 8 Moretti, F. “Network Theory, Plot Analysis.” NLR. May, 2011. 9 For each novel, a list of preliminary character names was generated using the Stanford NER, consisting of all proper names occurring 5 times. High error rates (8%-54%) required vetting by (1) checking entries against reference works on literary characters or (2) manually checking names in the body of the texts. Social networks are created for each novel by counting the number of paragraphs in which each pair of character names co-occurs. Edges are drawn for pairs with 3 co-occurrences and are weighted by the number of paragraphs in which characters co-occur, with darker lines indicating stronger, more frequent interactions. “Strong ties” are defined as co-occurrence in 20 paragraphs. 38 Intelligent Narrative Technologies: Papers from the 2012 AIIDE Workshop AAAI Technical Report WS-12-14