Frame Semantic Tree Kernels for Social Network Extraction from Text

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 211–219, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Frame Semantic Tree Kernels for Social Network Extraction from Text Apoorv Agarwal Dept. of Computer Science Columbia University New York, NY, USA Sriramkumar Balasubramanian Dept. of Computer Science Columbia University New York, NY, USA Anup Kotalwar Microsoft, Inc. Redmonad, WA, USA Jiehan Zheng The Peddie School Hightstown, NJ, USA Owen Rambow CCLS Columbia University New York, NY, USA apoorv@cs.columbia.edu Abstract In this paper, we present work on ex- tracting social networks from unstructured text. We introduce novel features de- rived from semantic annotations based on FrameNet. We also introduce novel se- mantic tree kernels that help us improve the performance of the best reported sys- tem on social event detection and classi- fication by a statistically significant mar- gin. We show results for combining the models for the two aforementioned sub- tasks into the overall task of social net- work extraction. We show that a combina- tion of features from all three levels of ab- stractions (lexical, syntactic and semantic) are required to achieve the best performing system. 1 Introduction Social network extraction from text has recently been gaining a considerable amount of attention (Agarwal and Rambow, 2010; Elson et al., 2010; Agarwal et al., 2013a; Agarwal et al., 2013b; He et al., 2013). One of the reason for this attention, we believe, is that being able to extract social net- works from unstructured text may provide a pow- erful new tool for historians, political scientists, scholars of literature, and journalists to analyze large collections of texts around entities and their interactions. The tool would allow researchers to quickly extract networks and assess their size, na- ture, and cohesiveness, a task that would otherwise be impossible with corpora numbering millions of documents. It would also make it possible to make falsifiable claims about these networks, bringing the experimental method to disciplines like his- tory, where it is still relatively rare. In our previous work (Agarwal et al., 2010), we proposed a definition of a network based on interactions: nodes are entities and links are so- cial events. We defined two broad types of links: one-directional links (one person thinking about or talking about another person) and bi-directional links (two people having a conversation, a meet- ing, etc.). For example, in the following sen- tence, we would add two links to the network: a one-directional link between Toujan Faisal and the committee, triggered by the word said (be- cause Toujan is talking about the committee) and a bi-directional link between the same entities trig- gered by the word informed (a mutual interaction). (1) [Toujan Faisal], 54, said [she] was informed of the refusal by an [Interior Ministry com- mittee] overseeing election preparations. In this paper, we extract networks using the aforementioned definition of social networks. We introduce and add tree kernel representations and features derived from frame-semantic parses to our previously proposed system. Our …

Content truncated to 500 words