Domain Adaptation from User-level Facebook Models to County-level Twitter Predictions

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 764–773, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Domain Adaptation from User-level Facebook Models to County-level Twitter Predictions Daniel Rieman Positive Psychology Center University of Pennsylvania rieman@seas.upenn.edu Kokil Jaidka Positive Psychology Center University of Pennsylvania jaidka@sas.upenn.edu H. Andrew Schwartz Computer and Information Science Stony Brook University has@cs.stonybrook.edu Lyle Ungar Computer and Information Science University of Pennsylvania ungar@cis.upenn.edu Abstract Several studies have demonstrated how language models of user attributes, such as personality, can be built by using the Facebook language of social media users in conjunction with their responses to psychology questionnaires. It is challenging to ap- ply these models to make general predictions about attributes of communities, such as personality dis- tributions across US counties, because it requires 1. the potentially inavailability of the original training data because of privacy and ethical regulations, 2. adapting Facebook language models to Twitter lan- guage without retraining the model, and 3. adapt- ing from users to county-level collections of tweets. We propose a two-step algorithm, Target Side Do- main Adaptation (TSDA) for such domain adapta- tion when no labeled Twitter/county data is avail- able. TSDA corrects for the different word distri- butions between Facebook and Twitter and for the varying word distributions across counties by ad- justing target side word frequencies; no changes to the trained model are made. In the case of pre- dicting the Big Five county-level personality traits, TSDA outperforms a state-of-the-art domain adap- tation method, gives county-level predictions that have fewer extreme outliers, higher year-to-year stability, and higher correlation with county-level outcomes. 1 Introduction Social media platforms offer an effective– and widely used– platform for administering surveys to individuals to measure their personality, socioe- conomic status, mental and physical well-being, and political orientation, which can then be com- bined with user posts to build language-based pre- dictive models of user attributes, traits and behav- iors. As compared to surveys, language models can be used to assess personality and well-being across communities of the U.S, at a scale not eas- ily achieved by surveys (Eichstaedt et al., 2015). In comparison, Twitter is a more effective tool to mine geographic trends from language, since tweets are publicly accessible, and one in five tweets can be mapped to the county from which they were sent (Schwartz et al., 2013b). However, to our knowledge, there are no tweet-based models of personality which are comparable in accuracy to the Facebook language models of personality. Figure 1: Predictions for county-level openness to expe- rience created by applying a user-level Facebook model for openness on TSDA adjusted Twitter county data Personality, as measured by the “Big Five” of openness, conscientiousness, extraversion, agree- ableness and neuroticism is known to vary region- ally worldwide (Rentfrow et al., 2013) and to cluster geographically in the United States (Rent- frow et al., 2013; Florida, 2002). In this paper, we wish to infer the regional variations of the Big Five Personality traits across the United …

Content truncated to 500 words