Predicting Human Activities from User-Generated Content

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2572–2582 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2572 Predicting Human Activities from User-Generated Content Steven R. Wilson and Rada Mihalcea University of Michigan {steverw,mihalcea}@umich.edu Abstract The activities we do are linked to our inter- ests, personality, political preferences, and de- cisions we make about the future. In this pa- per, we explore the task of predicting human activities from user-generated content. We collect a dataset containing instances of social media users writing about a range of everyday activities. We then use a state-of-the-art sen- tence embedding framework tailored to recog- nize the semantics of human activities and per- form an automatic clustering of these activi- ties. We train a neural network model to make predictions about which clusters contain activ- ities that were performed by a given user based on the text of their previous posts and self- description. Additionally, we explore the de- gree to which incorporating inferred user traits into our model helps with this prediction task. 1 Introduction What a person does says a lot about who they are. Information about the types of activities that a person engages in can provide insights about their interests (Goecks and Shavlik, 2000), per- sonality (Ajzen, 1987), physical health (Bouchard et al., 2018), the activities that they are likely to do in the future (Ouellette and Wood, 1998), and other psychological phenomena like personal val- ues (Rokeach, 1973). For example, it has been shown that university students who exhibit traits of interpersonal affect and self-esteem are more likely to attend parties (Paunonen and Ashton, 2001), and those that value stimulation are likely to watch movies that can be categorized as thrillers (Bardi and Schwartz, 2003). Several studies have applied computational ap- proaches to the understanding and modeling of hu- man behavior at scale (Yin et al., 2014) and in real time (Wang et al., 2015). However, this pre- vious work has mainly relied on specific devices or platforms that require structured definitions of behaviors to be measured. While this leads to an accurate understanding of the types of activities being done by the involved users, these methods capture a relatively narrow set of behaviors com- pared to the huge range of things that people do on a day-to-day basis. On the other hand, publicly available social media data provide us with infor- mation about an extremely rich and diverse set of human activities, but the data are rarely structured or categorized, and they mostly exist in the form of natural language. Recently, however, natural language processing research has provided several examples of methodologies for extracting and rep- resenting human activities from text (Fast et al., 2016; Wilson and Mihalcea, 2017) and even mul- timodal data (Agrawal et al., 2016). In this paper, we explore the task of predict- ing human activities from user-generated text data, which will allow us to gain a deeper understand- ing of the kinds of everyday …

Content truncated to 500 words