Best Practices in the Creation and Use of Emotion Lexicons

Findings of the Association for Computational Linguistics: EACL 2023, pages 1825–1836 May 2-6, 2023 ©2023 Association for Computational Linguistics Best Practices in the Creation and Use of Emotion Lexicons Saif M. Mohammad National Research Council Canada saif.mohammad@nrc-cnrc.gc.ca Abstract Words play a central role in how we express ourselves. Lexicons of word–emotion associa- tions are widely used in research and real-world applications for sentiment analysis, tracking emotions associated with products and policies, studying health disorders, tracking emotional arcs of stories, and so on. However, inappro- priate and incorrect use of these lexicons can lead to not just sub-optimal results, but also inferences that are directly harmful to people. This paper brings together ideas from Affec- tive Computing and AI Ethics to present, some of the practical and ethical considerations in- volved in the creation and use of emotion lexi- cons — best practices. The goal is to provide a comprehensive set of relevant considerations, so that readers (especially those new to work with emotions) can find relevant information in one place. We hope this work will facilitate more thoughtfulness when one is deciding on what emotions to work on, how to create an emotion lexicon, how to use an emotion lexi- con, how to draw meaningful inferences, and how to judge success. 1 Introduction Words often convey affect (emotions, sentiment, feelings, and attitudes); either explicitly through their core meaning (denotation) or implicitly through connotation. For example, dejected de- notes sadness. On the other hand, failure simply connotes sadness. Either through denotation or con- notation, both words are associated with sadness. A compilation of such associations is referred to as a word–affect association lexicon (aka emotion lexicon).1 An entry in a lexicon usually includes a word, an emotion category or affect dimension (e.g., joy, fear, valence, arousal, etc.), and a score indicating association (or strength of association). 1This includes sentiment lexicons that capture valence (as- sociation with the positive–negative dimension) and other lexica that capture affect-related phenomena. Examples of emotion lexicons include the Gen- eral Inquirer (Stone et al., 1966), ANEW (Nielsen, 2011; Bradley and Lang, 1999), LIWC (Pen- nebaker et al., 2001), Pittsburgh Subjectivity Lex- icon (Wilson et al., 2005), NRC Emotion Lexicon (Mohammad and Turney, 2010, 2013), and the NRC Valence, Arousal, and Dominance (VAD) Lex- icon (Mohammad, 2018). These were all created by manual annotation (either by experts or crowd- sourced). There also exist lexicons that were gen- erated automatically from large text corpora using statistical and/or machine learning algorithms; e.g., WordNet Affect (Strapparava et al., 2004), Senti- WordNet (SWN) (Baccianella et al., 2010). Emotion lexicons have a wide range of applica- tions in commerce, public health, and research (in NLP, Psychology, Social Sciences, Digital Human- ities, etc.). Some notable examples include: track- ing brand and product perception via social media posts, tracking support for controversial issues and policies, tracking buy-in for non-pharmaceutical health measures such as social distancing during a pandemic, literary analysis, and developing more natural dialogue systems. The lexicons can be used on their own or …

Content truncated to 500 words