You Shall Know a User by the Company It Keeps: Dynamic Representations for Social Media Users in NLP

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4707–4717, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4707 You Shall Know a User by the Company It Keeps: Dynamic Representations for Social Media Users in NLP Marco Del Tredici University of Amsterdam m.deltredici@uva.nl Diego Marcheggiani∗ Amazon marchegg@amazon.es Sabine Schulte im Walde University of Stuttgart schulte@ims.uni-stuttgart.de Raquel Fern´andez University of Amsterdam raquel.fernandez@uva.nl Abstract Information about individuals can help to bet- ter understand what they say, particularly in social media where texts are short. Current ap- proaches to modelling social media users pay attention to their social connections, but ex- ploit this information in a static way, treating all connections uniformly. This ignores the fact, well known in sociolinguistics, that an in- dividual may be part of several communities which are not equally relevant in all commu- nicative situations. We present a model based on Graph Attention Networks that captures this observation. It dynamically explores the social graph of a user, computes a user repre- sentation given the most relevant connections for a target task, and combines it with linguis- tic information to make a prediction. We apply our model to three different tasks, evaluate it against alternative models, and analyse the re- sults extensively, showing that it significantly outperforms other current methods. 1 Introduction The idea that extra-linguistic information about speakers can help language understanding has re- cently gained traction in NLP. Several studies have successfully exploited social information to clas- sify user-generated language in downstream tasks such as sentiment analysis (Yang and Eisenstein, 2017), abusive speech identification (Mishra et al., 2018) and sarcasm detection (Hazarika et al., 2018; Wallace et al., 2016). The underlying goal is to capture the sociological phenomenon of ho- mophily (McPherson et al., 2001) – i.e., people’s tendency to group together with others they share ideas, beliefs, and practices with – and to exploit it jointly with linguistic information to obtain richer text representations. In this paper, we advance ∗Research conducted when the author was at the Univer- sity of Amsterdam. this line of research. In particular, we address a common shortcoming in current models by using state-of-the-art graph neural networks to encode and leverage homophily relations. Most current models represent speakers as mul- tidimensional vectors derived by aggregating in- formation about all known social relations of a given individual in a uniform way. A major limit of this approach is that it does not take into ac- count the well known sociolinguistic observation that speakers typically belong to several commu- nities at once, in the sense of ‘communities of practice’ (Eckert and McConnell-Ginet, 1992) de- noting an aggregate of people defined by a com- mon engagement, such as supporters of a politi- cal party or fans of a TV show. Membership to these communities has different relevance in dif- ferent situations. For example, consider an indi- vidual who is both part of a supporters group of the …

Content truncated to 500 words