CHIRON: Rich Character Representations in Long-Form Narratives

TL;DR

CHIRON addresses the challenge of representing complex, long-form characters by coupling a Generation Module that constructs structured, writer-inspired character sheets with a Validation Module that uses automated reasoning and a domain-specific entailment model to prune unfaithful statements. The approach yields robust character representations across four categories (Dialogue, Physical/Personality, Knowledge, Goals) and improves downstream tasks such as masked-character prediction by $11.6\%$ over summary baselines. A density metric derived from CHIRON correlates highly with human judgments of character-centricity, enabling automatic analysis across varied story datasets. Together, these contributions advance character-centric narrative analysis and generation, with practical impact for downstream Story AI systems and literary understanding.

Abstract

Characters are integral to long-form narratives, but are poorly understood by existing story analysis and generation systems. While prior work has simplified characters via graph-based methods and brief character descriptions, we aim to better tackle the problem of representing complex characters by taking inspiration from advice given to professional writers. We propose CHIRON, a new `character sheet' based representation that organizes and filters textual information about characters. We construct CHIRON sheets in two steps: a Generation Module that prompts an LLM for character information via question-answering and a Validation Module that uses automated reasoning and a domain-specific entailment model to eliminate false facts about a character. We validate CHIRON via the downstream task of masked-character prediction, where our experiments show CHIRON is better and more flexible than comparable summary-based baselines. We also show that metrics derived from CHIRON can be used to automatically infer character-centricity in stories, and that these metrics align with human judgments.

Figures (4)

• Figure 1: Overview of CHIRON's Generation and Validation Modules. For each CHIRON category, our Generation Module takes a character and story snippet and uses a pretrained LLM to generate statements about the character. Our Validation Module then passes these statements through reasoning steps and a trained entailment classifier to determine if they are true and useful. We find this methodology produces more accurate and useful character-representations for downstream tasks and analysis. We use Mistral 7B-Instruct v0.2 unless otherwise stated.
• Figure 2: Guidelines and consent for task shown to Prolific annotators.
• Figure 3: Instructions for annotation task shown to Prolific annotators, including one of the provided examples.
• Figure 4: Example of annotation task shown to Prolific annotators.