Shaping Event Backstories to Estimate Potential Emotion Contexts
Johannes Schäfer, Roman Klinger · 2025
Read Shaping Event Backstories to Estimate Potential Emotion Contexts by Johannes Schäfer, Roman Klinger (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-09954.txt.
Introduction
Johannes Schäfer and Roman Klinger build a method to add plausible backstories to ambiguous emotion texts, in work from 2025. Emotion labels can be unclear when context is missing, not only when readers differ. Their method generates short narratives that fill gaps in a way a reader might, using techniques from short story generation.
The resulting English dataset, Emotional BackStories, supports a systematic study of emotion analysis with added context. Human and automatic checks found that the stories do clarify specific emotions. Relief and sadness gained most from the added context, while joy did not need it. The work offers a new way to study how much context emotion judgments depend on.
Catalog record
| CREATOR | Johannes Schäfer, Roman Klinger |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 97 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-09954
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Shaping Event Backstories to Estimate Potential Emotion Contexts Johannes Schäfer, Roman Klinger Abstract Emotion analysis is an inherently ambiguous task. Previous work studied annotator properties to explain disagreement, but this overlooks the possibility that ambiguity may stem from missing information about the context of events. In this paper, we propose a novel approach that adds reasonable contexts to event descriptions, which may better explain a particular situation. Our goal is to understand whether these enriched contexts enable human annotators to annotate emotions more reliably. We disambiguate a target event description by automatically generating multiple event chains conditioned on differing emotions. By combining techniques from short story generation in various settings, we achieve coherent narratives that result in a specialized dataset for the first comprehensive and systematic examination of contextualized emotion analysis. Through automatic and human evaluation, we find that contextual narratives enhance the interpretation of specific emotions and support annotators in producing more consistent annotations. 1 Introduction Guilt: A close friend entrusted me with setting up the sound system for a charity event. I faced a daunting task, as the previous technician had left the equipment in disarray. With time running out, I took a shortcut to meet the deadline, skipping some crucial safety checks. Just before the event started, my friend reminded me of its significance and the many people counting on its success. The loudspeaker suddenly malfunctioned and went silent. Relief: I was tasked with giving a presentation to a large crowd. The sound system malfunctioned, amplifying my voice to an ear-piercing level. The sound technician ignored the problem and chatted with someone. The audience covered their ears and looked at me with discomfort. The loudspeaker suddenly malfunctioned and went silent. Fear: I arrived at a remote wilderness survival training camp, where the instructors emphasized the importance of following loudspeaker instructions for safety. The instructors warned us about the toxic waste site nearby and explained that the loudspeaker would alert us to any changes in air quality. During the first exercise, I struggled to navigate the challenging terrain, but the loudspeaker provided crucial guidance, helping me stay on track. I completed a difficult obstacle course, relying heavily on the loudspeaker’s instructions to avoid hazards and find the safest route. The loudspeaker suddenly malfunctioned and went silent. Pride: I spent the entire morning upgrading the sound system with a new backup system to prevent technical issues. The event organizer informed me that the conference was running 30 minutes behind schedule, giving me extra time to test the new backup system. I used the extra time to run a series of tests on the sound system, trying to simulate potential failures. The keynote speaker began to talk, and the sound system was working flawlessly, but I was still waiting for a real test of the new backup system. The loudspeaker suddenly malfunctioned and went silent. Table 1 : Examples of stories evoking different emotions. In analyzing event descriptions, subtle variations in context can significantly influence emotion responses. For instance, given the following text from the dataset we present in this paper 1 1 1 We provide our annotated dataset and code on https://www.uni-bamberg.de/en/nlproc/resources/emotional-backstories/ . : “The loudspeaker suddenly malfunctioned and went silent.”, a prevalent interpretation may be that this event causes surprise. However, the context of the event may change the distribution of possibly experienced emotions as shown in Table 1 . In the case of guilt, the added narrative establishes a profound sense of personal responsibility and the anticipated negative consequences of actions, which are then triggered by the event. Relief emerges from including a backstory that portrays a situation initially fraught with anxiety and possibly negative consequence beyond one’s control, suggesting that these are resolved by the event. Fear is framed by a narrative that expresses the unpredictability of a challenging circumstance, emphasizing a sense of uncertainty and the immense effort required to navigate the unknown, reinforcing an unpleasant loss of control based on the event. Lastly, pride is evoked by incorporating a backstory that highlights proactive efforts and successful task management, illustrating how these accomplishments will be recognized by others as a result of the event. These examples demonstrate that backstories elicit cognitive appraisals for each response, which lead to emotion interpretation differences. While there have been studies on emotion classification under the paradigm of perspectivism, we are not aware of any work that studies the importance of context to disambiguate emotion analysis in a controlled manner. 2 2 2 It should be noted however, that emotion recognition in conversations typically does consider context, but not to disambiguate otherwise unclear utterances in a controlled manner ( Hu et al. 2021 , i.a.) . Ambiguity in the interpretation of event descriptions viewed in isolation can stem from missing essential information, which can also lead to uncertainty in emotion responses. Furthermore, not all event descriptions elicit distinct emotions; some are devoid of specific emotion, and thus can be filled only through narratives. We approach disambiguation by generating different contexts as sequences of preceding events, thereby examining how various backstories can have different effects on the emotion interpretation of that event. Our proposed systems generate event chains by prompting a large language model (LLM) in different settings, with the goal to understand whether the contexts generated with these methods enable humans to annotate emotions more reliably. Our paper is guided by the following research questions: 1. Does context generation through iterative prompting with story planning enhance narrative coherence? 2. Do generated contexts enhance clarity in human annotation of emotions in events? 3. Can systems recognize the contextual disambiguation in emotion analysis of events? We construct a dataset which lays the foundation for a comprehensive analysis of how incorporating contextual information can shape emotion interpretations of events. In Section 2 we review related work, in Section 3 we present our methods and in Section 4 we analyze results of our experiments. 2 Related Work The novel task of generating backstories that evoke specific emotions in events intersects with multiple research fields, which we review in this section. 2.1 Contextual Influences on Event Interpretations Context significantly influences interpretations of text. In a study of event chains, Das et al. 2011 explore emotion dynamics with a focus on temporal ordering. Mostafazadeh et al. 2016 introduce the ROCstory dataset to facilitate story completion tasks. Park et al. 2020 perform sentence-level emotion analysis on this dataset. The GLUCOSE dataset by Mostafazadeh et al. 2020 enhances understanding of complex dependencies between events based on commonsense inferences; however, it does not specifically address the emotions being evoked in these contexts. 2.2 Emotion Categorization in Context Emotion analysis seeks to identify emotional states elicited in readers or authors of texts. The foundational frameworks by Ekman 1972 and Plutchik 2001 provide categorization methodologies. Mohammad 2012 created emotion-labeled social media corpora, while Troiano et al. 2023 emphasized implicit emotion cues linked to appraisal theories. Their analysis of event descriptions without context highlights a gap our study addresses by focusing on contextual influences. Notably, Etienne et al. 2022 emphasize the intricacies of emotion expression across diverse texts, and Étienne et al. 2024 predict various modes of emotion expression, emphasizing the complexity that this context brings to emotion annotation. Chatterjee et al.