Automatic Extraction of Social Networks from Literary Text: A Case Study on Alice in Wonderland

International Joint Conference on Natural Language Processing, pages 1202–1208, Nagoya, Japan, 14-18 October 2013. Automatic Extraction of Social Networks from Literary Text: A Case Study on Alice in Wonderland Apoorv Agarwal Dept. of Computer Science Columbia University New York, NY, U.S.A. apoorv@cs.columbia.edu Anup Kotalwar Microsoft, Inc. Redmond, WA, U.S.A. ankotalw@microsoft.com Owen Rambow CCLS Columbia University New York, NY, U.S.A. rambow@ccls.columbia.edu Abstract In this paper we present results for two tasks: social event detection and social network extraction from a literary text, Al- ice in Wonderland. For the first task, our system trained on a news corpus using tree kernels and support vector machines beats the baseline systems by a statisti- cally significant margin. Using this sys- tem we extract a social network from Al- ice in Wonderland. We show that while we achieve an F-measure of about 61% on social event detection, our extracted un- weighted network is not statistically dis- tinguishable from the un-weighted gold network according to popularly used net- work measures. 1 Introduction Social network analysis affects a wide range of academic disciplines and practical applica- tions: psychology (Seidman, 1985; Koehly and Shivy, 1998), anthropology (Sanjek, 1974; John- son, 1994; Hage and Harary, 1983), political sci- ence (Knoke, 1990; Brandes et al., 2001), liter- ary theory (Moretti, 2005), management (Tichy et al., 1979; Cross et al., 2001; Borgatti and Cross, 2003), and crime prevention and intelli- gence (Sparrow, 1991). In the past, social net- works were constructed through interviews, sur- veys and experiments. With the advent of the internet and online social networks, researchers have started constructing networks using meta- data that reflects interactions, such as self-declared friendship linkages, sender-receiver email link- ages, comments on a common blog-post, etc. However, these methodologies of creating social networks ignore a vastly rich network expressed in the unstructured text of such sources. More- over, many rich sources of social networks remain untouched simply because there is no meta-data associated with them (literary texts, news stories, historical texts). There have been recent efforts to extract social networks from text by mining in- teractions between people expressed linguistically in unstructured text (Elson et al., 2010; He et al., 2013). However, these approaches are restricted to extracting interactions signaled by quoted speech. In this paper, we present results for extracting a social network from Alice in Wonderland that is not restricted to interactions signaled by quoted speech. We define a social network for a fictional text as follows: nodes are characters and links are social events. Two nodes in the network are con- nected if the characters engage in a social event. We introduced the notion of social events in our previous work (Agarwal et al., 2010), in which we presented our annotation scheme for annotating social events on the Automatic Content Extraction (ACE-2005) corpus. We presented a preliminary system for social event detection and classifica- tion in Agarwal and Rambow (2010). This system was trained and tested only on the ACE-2005 cor- pus. A priori, it is unclear if a system trained …

Content truncated to 500 words