2023.bigpicture-1.1
The paper surveys the field of event-centric emotion analysis in natural language processing, contrasting two dominant strands: (1) emotion role labeling, which treats emotions as events and extracts spans for experiencers, cues, and causes, and (2) appraisal-based approaches, which model the cognitive evaluation of events that give rise to emotions. It begins with a review of psychological emotion theories—categorical, dimensional, and appraisal—highlighting the consensus that emotions are caused by events and have multiple subcomponents. The authors then map a wide range of related NLP tasks (sentiment analysis, personality profiling, etc.) and use‑cases (social media, news, literature, clinical data). The survey consolidates recent work on (a) interpreting events to predict emotions via appraisal variables, including traditional feature‑based systems, modern pretrained language models, and newly created appraisal‑annotated corpora; and (b) structured emotion role labeling, covering cue detection, stimulus detection, and full role annotation efforts across languages and domains. It discusses limited attempts to bridge the two perspectives, showing that role information can aid emotion classification and that appraisal variables can enrich role labeling, but full integration remains underexplored. The paper concludes with an extensive list of open research challenges: full emotion graph prediction, joint modeling of roles and appraisals, incorporation of dimensional representations, cross‑corpus and zero‑shot robustness, event‑chain interpretation, perspectivism, multimodal and multilingual extensions, and deeper integration of psychological theories. The authors position the work as a roadmap for researchers aiming to develop more comprehensive, theory‑grounded emotion analysis systems.
Topics
Dual Event Perspectives: Emotions as Events and Emotions Caused by Events
Introduces the two central viewpoints that frame the survey: (1) treating emotions themselves as events for role labeling, and (2) viewing emotions as outcomes of event appraisal, establishing the conceptual bridge between them.
Psychological Foundations: Categorical, Dimensional, and Appraisal Theories
Summarizes key emotion theories from psychology—basic emotion categories (Ekman, Plutchik), valence/arousal dimensions, and appraisal models (Smith/Ellsworth, OCC)—and explains why they matter for NLP modeling.
Task Landscape and Domain Use‑Cases for Emotion Analysis
Maps the spectrum of NLP tasks (emotion classification, role labeling, sentiment analysis, personality profiling) and illustrates domain‑specific requirements (social media, news, literature, clinical texts).
Appraisal‑Based Emotion Classification: From Feature‑Based Systems to Pretrained Models
Reviews the evolution of appraisal‑driven approaches, covering early lexicon‑based methods, rule‑based extraction of appraisal variables, and recent fine‑tuned transformer models trained on appraisal‑annotated corpora (e.g., enVENT, APPReddit).
Emotion Role Labeling: Structured Extraction of Cues, Stimuli, and Experiencers
Details the corpus and modeling efforts for role labeling, including cue detection, stimulus detection, full role annotation (REMAN, REMAN, GoodNewsEveryone), and the challenges of span prediction and relation detection.
Bridging Role Labeling and Appraisal: Early Integration Attempts
Discusses the few studies that combine role information with appraisal variables, showing mutual benefits but also highlighting the lack of full graph‑level integration.
Open Research Directions: Towards Comprehensive Emotion Graphs
Enumerates future challenges such as full emotion graph prediction, joint modeling of roles and appraisals, dimensional representations, cross‑corpus/zero‑shot robustness, event‑chain reasoning, perspectivism, multimodal and multilingual extensions, and deeper incorporation of psychological theories.