That’s so cute!: The CARE Dataset for Affective Response Detection

Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 50 - 69 December 7-8, 2022 ©2022 Association for Computational Linguistics “That’s so cute!”: The CARE Dataset for Affective Response Detection Jane Dwivedi-Yu Meta AI janeyu@meta.com Alon Y. Halevy Meta AI ayh@meta.com Abstract Social media plays an increasing role in our communication with friends and family, and in our consumption of entertainment and in- formation. Hence, to design effective ranking functions for posts on social media, it would be useful to predict the affective responses of a post (e.g., whether it is likely to elicit feelings of entertainment, inspiration, or anger). Similar to work on emotion detection (which focuses on the affect of the publisher of the post), the traditional approach to recognizing affective re- sponse would involve an expensive investment in human annotation of training data. We create and publicly release CAREdb, a dataset of 230k social media post annotations according to seven affective responses using the Common Affective Response Expression (CARE) method. The CARE method is a means of leveraging the signal that is present in comments that are posted in response to a post, providing high-precision evidence about the affective response to the post without human annotation. Unlike human annotation, the anno- tation process we describe here can be iterated upon to expand the coverage of the method, particularly for new affective responses. We present experiments that demonstrate that the CARE annotations compare favorably with crowdsourced annotations. Finally, we use CAREdb to train competitive BERT-based mod- els for predicting affective response as well as emotion detection, demonstrating the utility of the dataset for related tasks. 1 Introduction Social media and other online media platforms have become a common means of not only inter- acting and connecting with others, but also find- ing interesting, informing, and entertaining content. Users of those platforms depend on the ranking sys- tems of the recommendation systems to show them information they will be most interested in and to safeguard them against unfavorable experiences. Figure 1: Overview of the CARE Method (pseudo-code in Appendix, Algorithm 1). The top half of the figure (steps 1–3) shows how the affective response to a post is computed by aggregating the expressed affects in comments from users viewing the post. The bottom half of the figure (steps A–C) shows how we expand the collection of CARE patterns and the lexicon based on labels that have been obtained from prior iterations. Towards this end, a key technical problem is to predict the affective response that a user may have when they see a post. Some affective responses can be described by emotions (e.g., angry, joyful), and others may be described more as experiences (e.g., entertained, inspired). Predicting affective response differs from emotion detection in that the latter focuses on the emotions expressed by the publisher of the post (referred to as the publisher 50 affect in Chen et al. (2014)) and not on the viewer of the content. While the publisher’s emotion may be relevant to …

Content truncated to 500 words