User Classification with Multiple Textual Perspectives

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2112–2121, Osaka, Japan, December 11-17 2016. User Classification with Multiple Textual Perspectives Dong Zhang, Shoushan Li, Hongling Wang, Guodong Zhou Natural Language Processing Lab School of Computer Science and Technology, Soochow University, China dzhang@stu.suda.edu.cn, lishoushan@suda.edu.cn hlwang@suda.edu.cn, gdzhou@suda.edu.cn Abstract Textual information is of critical importance for automatic user classification in social media. However, most previous studies model textual features in a single perspective while the text in a user homepage typically possesses different styles of text, such as original message and comment from others. In this paper, we propose a novel approach, namely ensemble LSTM, to user classification by incorporating multiple textual perspectives. Specifically, our approach first learns a LSTM representation with a LSTM recurrent neural network and then presents a joint learning method to integrating all naturally- divided textual perspectives. Empirical studies on two basic user classification tasks, i.e., gender classification and age classification, demonstrate the effectiveness of the proposed approach to user classification with multiple textual perspectives. 1 Introduction User attribute classification, also namely user classification for short, is a task which aims to leverage user-generated content to automatically predict user’s attributes, such as gender (Wang et al., 2015), age (Rao et al., 2010; Sap et al., 2014) and location (Cheng et al., 2010). Recently, the growth of online social networks provides the opportunity to perform user classification in a broader context (Bollen et al., 2011; Sadilek et al., 2012; Lampos and Cristianini, 2010; Zamal et al., 2012). Basically, user classification is a fundamental task not only in sociolinguistic studies, but also in many real applications, such as recommender systems, and online advertising (O’Connor et al., 2010; Preotiuc- Pietro et al, 2015). Text style User A Gender: female User B Gender: male Original Message “Just bought the lipstick, look beautiful?” “The first day, hard work.” Retweeted Message “Seaweed mask, it is so remarkably efficient.” “Love her, take her to see the sea.” Comment from others “Sister, you’re so pretty!” “Go to see my latest message” Comment to others “Thanks.” “Sister, you’re so pretty!” Table 1: Some examples of different text styles in two users’ homepages in a social media Currently, machine learning approaches have dominated the research on user classification where statistic classifiers are learned with labeled data and various kinds of features, such as textual features, This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http://creativecommo ns.org/licenses/by/4.0/  Corresponding author 2112 behavior features, and social connection features (Preotiuc et al., 2015, Lampos et al., 2016). Among these features, textual features are most popular and they are good clues to infer the user attributes (Zhu et al., 2015; Li et al., 2015). For example, in Table 1, User A publishes a text “Just bought the lipstick, look beautiful?” which could be used to infer the user to be a female since females are more likely to buy a lipstick. However, user-generated text sometimes possesses different styles, especially in social media. …

Content truncated to 500 words