Human Centered NLP with User-Factor Adaptation

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1146–1155 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Human Centered NLP with User-Factor Adaptation Veronica E. Lynn, Youngseo Son, Vivek Kulkarni Niranjan Balasubramanian and H. Andrew Schwartz Stony Brook University Stony Brook, NY {velynn, yson, vvkulkarni, niranjan, has}@cs.stonybrook.edu Abstract We pose the general task of user-factor adaptation — adapting supervised learn- ing models to real-valued user factors in- ferred from a background of their lan- guage, reflecting the idea that a piece of text should be understood within the con- text of the user that wrote it. We intro- duce a continuous adaptation technique, suited for real-valued user factors that are common in social science and bringing us closer to personalized NLP, adapting to each user uniquely. We apply this tech- nique with known user factors including age, gender, and personality traits, as well as latent factors, evaluating over five tasks: POS tagging, PP-attachment, sentiment analysis, sarcasm detection, and stance de- tection. Adaptation provides statistically significant benefits for 3 of the 5 tasks: up to +1.2 points for PP-attachment, +3.4 points for sarcasm, and +3.0 points for stance. 1 Introduction Language use is personal. Knowing who wrote a piece of text can help to better understand it. For instance, knowing the age and gender groups of authors has been shown to improve document classification (Hovy, 2015) and sentiment ana- lysis (Volkova et al., 2013). However, putting people into discrete groups (e.g. age groups, binary gender) often relies on arbitrary boundaries which may not correspond to meaningful changes in language use. A wealth of psychological research suggests people should not be characterized as discrete types (or domains) but rather as mixtures of continuous factors (McCrae and Costa Jr., 1989; Ruscio and Ruscio, 2000; Widiger and Samuel, 2005). Here, we ask how one can adapt NLP models to real-valued human factors – continuous valued attributes that capture fine-grained differences be- tween users (e.g. real-valued age, continuous gen- der scores). We refer to this problem as user- factor adaptation, and investigate a solution to it in the context of social media, a genre where lan- guage is generated by a particularly diverse set of users (Duggan and Smith, 2013). Importantly, user-factor adaptation brings us closer to person- alized NLP in that with real-valued factors we can now adapt uniquely for each user. Our approach composes user factor information with the linguistic features, similar to feature aug- mentation (Daum´e III, 2007), a widely used do- main adaptation technique which allows for easy integration with most feature-based learning mod- els. Since relevant user information often is not explicitly available, we use a background of tweets from the user to infer user factors. We evaluate our approach over five tasks — POS tagging, PP- attachment, sentiment analysis, sarcasm detection, and stance detection — and with a variety of in- ferred user factors including (a) known factors: age, gender, and personality traits, as well as (b) latent factors derived from past …

Content truncated to 500 words