Contrastive Learning with Narrative Twins for Modeling Story Salience

TL;DR

The paper addresses modeling narrative salience by learning story embeddings through contrastive learning with narrative twins (same plot, different surface form) and distractors. It introduces a window-based, transformer-backed model trained with the InfoNCE objective, and evaluates four salience operations—deletion, shifting, disruption, and summarization—on short ROCStories and long Wikipedia plot summaries. Empirical results show that contrastively learned embeddings outperform masked language-model baselines, with summarization consistently providing the most reliable salience signal; dropout twins offer a practical alternative when explicit narrative twins are unavailable. The work advances the understanding of how to operationalize narratological salience in neural representations and suggests directions for scalable, multi-level narrative analysis in long-form texts.

Abstract

Understanding narratives requires identifying which events are most salient for a story's progression. We present a contrastive learning framework for modeling narrative salience that learns story embeddings from narrative twins: stories that share the same plot but differ in surface form. Our model is trained to distinguish a story from both its narrative twin and a distractor with similar surface features but different plot. Using the resulting embeddings, we evaluate four narratologically motivated operations for inferring salience (deletion, shifting, disruption, and summarization). Experiments on short narratives from the ROCStories corpus and longer Wikipedia plot summaries show that contrastively learned story embeddings outperform a masked-language-model baseline, and that summarization is the most reliable operation for identifying salient sentences. If narrative twins are not available, random dropout can be used to generate the twins from a single story. Effective distractors can be obtained either by prompting LLMs or, in long-form narratives, by using different parts of the same story.

Figures (3)

• Figure 1: Left: Overview of contrastive training for learning narrative embeddings. Inputs are an anchor narrative $\mathbf{x}$, its narrative twin $\mathbf{x^+}$, and a distractor $\mathbf{x^-}$. A shared encoder $f_\theta$ generates embeddings. The embeddings are optimized with a contrastive loss to be similar for twins and dissimilar to distractors (and in-batch negatives). Right: At inference, we score the salience of each sentence via four theoretically motivated operations.
• Figure 2: Validation performance on ROCStories throughout training. Curves show Spearman rank correlation (which has range [-1, 1]; random is 0.0; higher is better) between model-predicted salience and human rankings. All models are trained with the same data. In blue is a model trained with masked language modeling. In red is a model trained with contrastive learning and positives derived from applying different random dropout on the same input. Black line represents model trained with textual narrative twins.
• Figure 3: ROCStory model performance with increasing model size.