Modeling the Use of Graffiti Style Features to Signal Social Relations within a Multi-Domain Learning Paradigm

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 107–115, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Modeling the Use of Graffiti Style Features to Signal Social Relations within a Multi-Domain Learning Paradigm Mario Piergallini1, A. Seza Doğruöz2, Phani Gadde1, David Adamson1, Carolyn P. Rosé1,3 1Language Technologies Institute Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh PA, 15213 {mpiergal,pgadde, dadamson}@cs.cmu.edu 2Tilburg University, TSH, 5000 LE Tilburg, The Netherlands/ Language Technologies Institute, Carnegie Mellon University, 5000 Forbes Ave.,Pittsburgh PA 15213 a.s.dogruoz@gmail.com 3Human-Computer Interaction Institute Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh PA, 15213 cprose@cs.cmu.edu Abstract In this paper, we present a series of experiments in which we analyze the usage of graffiti style features for signaling personal gang identification in a large, online street gangs forum, with an accuracy as high as 83% at the gang alliance level and 72% for the specific gang. We then build on that result in predicting how members of different gangs signal the relationship between their gangs within threads where they are interacting with one another, with a predictive accuracy as high as 66% at this thread composition prediction task. Our work demonstrates how graffiti style features signal social identity both in terms of personal group affiliation and between group alliances and oppositions. When we predict thread composition by modeling identity and relationship simultaneously using a multi-domain learning framework paired with a rich feature representation, we achieve significantly higher predictive accuracy than state-of-the-art baselines using one or the other in isolation. 1 Introduction Analysis of linguistic style in social media has grown in popularity over the past decade. Popular prediction problems within this space include gender classification (Argamon et al., 2003), age classification (Argamon et al., 2007), political affiliation classification (Jiang & Argamon, 2008), and sentiment analysis (Wiebe et al., 2004). From a sociolinguistic perspective, this work can be thought of as fitting within the area of machine learning approaches to the analysis of style (Biber & Conrad, 2009), perhaps as a counterpart to work by variationist sociolinguists in their effort to map out the space of language variation and its accompanying social interpretation (Labov, 2010; Eckert & Rickford, 2001). One aspiration of work in social media analysis is to contribute to this literature, but that requires that our models are interpretable. The contribution of this paper is an investigation into the ways in which stylistic features behave in the language of participants of a large online community for street gang members. We present a series of experiments that reveal new challenges in modeling stylistic variation with machine learning approaches. As we will argue, the challenge is achieving high predictive accuracy without sacrificing interpretability. Gang language is a type of sociolect that has so far not been the focus of modeling in the area of social media analysis. Nevertheless, we argue that the gangs forum we have selected as our data source provides a strategic source of data for exploring how social context influences stylistic language …

Content truncated to 500 words