Uncovering Surprising Event Boundaries in Narratives

Abstract

When reading stories, people can naturally identify sentences in which a new event starts, i.e., event boundaries, using their knowledge of how events typically unfold, but a computational model to detect event boundaries is not yet available. We characterize and detect sentences with expected or surprising event boundaries in an annotated corpus of short diary-like stories, using a model that combines commonsense knowledge and narrative flow features with a RoBERTa classifier. Our results show that, while commonsense and narrative features can help improve performance overall, detecting event boundaries that are more subjective remains challenging for our model. We also find that sentences marking surprising event boundaries are less likely to be causally related to the preceding sentence, but are more likely to express emotional reactions of story characters, compared to sentences with no event boundary.

Modeling Empathic Similarity in Personal Narratives

“…Most relevant to our work are recent advances in social and emotional commonsense reasoning using using language models. Specifically, prior methods have used finetuning of language models such as BERT (Devlin et al, 2019;Reimers and Gurevych, 2019) and GPT-2 (Radford et al) to model events and the emotional reactions caused by everyday events (Rashkin et al, , 2018Sap et al, 2019b;Bosselut et al, 2019;Wang et al, 2022;West et al, 2022;Mostafazadeh et al, 2020) as well as predicting empathy, condolence, or prosocial outcomes (Lahnala et al, 2022a;Kumano et al;Boukricha et al, 2013;Zhou and Jurgens, 2020;Bao et al, 2021). Understanding the emotional reactions elicited by events is a challenging task for many NLP systems, as it requires commonsense knowledge and extrapolation of meanings beyond the text alone.…”

Event Knowledge in Large Language Models: The Gap Between the Impossible and the Unlikely

“…The term “event” has different meanings across disciplines and can encompass both an individual action or a sequence of several actions (Zacks, 2020; Zacks, Speer, Swallow, Braver, & Reynolds, 2007; see Kuperberg, 2021 for discussion). Some research on event knowledge in LLMs, for example, asks whether LLMs trained on word‐in‐context prediction encode human‐like knowledge of event boundaries, investigating their capacity to replicate a fundamental aspect of human cognitive processing related to understanding sequential events in narratives (Kumar et al., 2022; Michelmann, Kumar, Norman, & Toneva, 2023; Wang, Jafarpour, & Sap, 2022).…”