Neural User Factor Adaptation for Text Classification: Learning to Generalize Across Author Demographics

Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM), pages 136–146 Minneapolis, June 6–7, 2019. c⃝2019 Association for Computational Linguistics 136 Neural User Factor Adaptation for Text Classification: Learning to Generalize Across Author Demographics Xiaolei Huang and Michael J. Paul Information Science University of Colorado Boulder, CO 80309, USA {xiaolei.huang,mpaul}@colorado.edu Abstract Language usage varies across different demo- graphic factors, such as gender, age, and ge- ographic location. However, most existing document classification methods ignore demo- graphic variability. In this study, we exam- ine empirically how text data can vary across four demographic factors: gender, age, coun- try, and region. We propose a multitask neural model to account for demographic variations via adversarial training. In experiments on four English-language social media datasets, we find that classification performance im- proves when adapting for user factors. 1 Introduction Different demographic groups can show sub- stantial linguistic variations, especially in online data (Goel et al., 2016; Johannsen et al., 2015). These variations can affect natural language pro- cessing models such as sentiment classifiers. For example, researchers found that women were more likely to use the word weakness in a positive way, while men were more likely to use the word in a negative expression (Volkova et al., 2013). Models for text classification, the automatic cat- egorization of documents into categories, typi- cally ignore attributes about the authors of the text. With the growing amount of text generated by users online, whose personal characteristics are highly variable, there has been increased atten- tion to how user demographics are associated with the text they write. Promising recent studies have shown that incorporating demographic factors can improve text classification (Volkova et al., 2013; Hovy, 2015; Yang and Eisenstein, 2017; Li et al., 2018). Lynn et al. (2017) refer to this idea as user factor adaptation and proposed to treat this as a domain adaptation problem in which demographic attributes constitute different domains. We extend this line of work in a number of ways: • We assemble and publish new datasets con- taining four demographic factors: gender, age, country, and US region. The demographic at- tributes are carefully inferred from profile in- formation that is separate from the text data. • We experiment with neural domain adaptation models (Ganin et al., 2016), which may pro- vide better performance than the simpler mod- els used in prior work on user factor adaptation. We also propose a new model using a multitask framework with adversarial training. • Our approach requires demographic attributes at training time but not at test time: we learn a single representation to be invariant to demo- graphic changes. This approach thus requires fewer resources than prior work. In this study, we treat adapting across the de- mographic factors as a domain work problem, in which we consider each demographic factor as a domain. We focus on four different de- mographic factors (gender, age, country, region) in four English-language social media datasets (Twitter, Amazon reviews, Yelp hotel reviews, and Yelp restaurant reviews), which contain …

Content truncated to 500 words