Entity-Centric Contextual Affective Analysis

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2550–2560 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2550 Entity-Centric Contextual Affective Analysis Anjalie Field Carnegie Mellon University anjalief@cs.cmu.edu Yulia Tsvetkov Carnegie Mellon University ytsvetko@cs.cmu.edu Abstract While contextualized word representations have improved state-of-the-art benchmarks in many NLP tasks, their potential usefulness for social-oriented tasks remains largely un- explored. We show how contextualized word embeddings can be used to capture affect di- mensions in portrayals of people. We evaluate our methodology quantitatively, on held-out affect lexicons, and qualitatively, through case examples. We find that contextualized word representations do encode meaningful affect information, but they are heavily biased to- wards their training data, which limits their usefulness to in-domain analyses. We ulti- mately use our method to examine differences in portrayals of men and women. 1 Introduction Pre-trained contextualized word embeddings (Pe- ters et al., 2018; Devlin et al., 2019; Radford et al., 2018) have become increasingly common in natu- ral language processing (NLP), improving state- of-the-art results in many standard NLP tasks. However, beyond standard tasks, NLP tools are also vital to more open-ended exploratory tasks, particularly in social science. How these types of tasks can benefit from pre-trained contextualized embeddings has not yet been explored. In this work, we show how to leverage these em- beddings to conduct entity-centric analyses, which broadly seek to address how entities are portrayed in narrative text (Bamman et al., 2013; Card et al., 2016). For instance, in the sentence “Batman ap- prehends the Joker”, a reader might infer that Bat- man is good, the Joker is evil, and Batman is more powerful than the Joker. Analyzing how people are portrayed in narratives is a key starting point to identifying stereotypes and bias (Joseph et al., 2017; Fast et al., 2016; Field et al., 2019). Existing methods for analyzing people portray- als take either an unsupervised approach (Bam- man et al., 2013), which requires large amounts of data and can be difficult to interpret, or rely on domain-specific knowledge (Fast et al., 2016; Wagner et al., 2015), which does not general- ize well to other hypotheses and data domains. Furthermore, most models are limited to discrete word-level features, whereas continuous-valued embeddings are typically more expressive. We in- troduce a novel approach to analyzing entities that maps contextualized embeddings to interpretable dimensions. Specifically, we propose using pre-trained em- beddings to extract affect information about tar- get entities. Social psychology research has identified 3 primary affect dimensions: Potency (strength/weakness of an identity), Valence (good- ness/badness of an identity), and Activity (active- ness/passiveness of an identity) (Osgood et al., 1957; Russell, 1980, 2003). We refer to these di- mensions as power, sentiment, and agency for consistency with prior work in NLP (Sap et al., 2017; Rashkin et al., 2016; Field et al., 2019). Thus, in the previous example, “Batman appre- hends the Joker”, we might associate Batman with high power, high sentiment, and high agency. While much literature in …

Content truncated to 500 words