"Show, Don't Tell: Uncovering Implicit Character Portrayal using LLMs"

LIIPA: Teaching machines to understand character development like human readers do.

This paper introduces LIIPA (LLMs for Inferring Implicit Portrayal for Character Analysis), a framework that uses LLMs to uncover hidden character traits in stories.

Unlike existing tools that rely on explicit descriptions, LIIPA analyzes actions and behaviors to reveal character portrayals across intellect, appearance, and power dimensions.

-----

https://arxiv.org/abs/2412.04576

🔍 Original Problem:

→ Current tools for analyzing fictional characters mainly use explicit textual indicators, missing subtle character traits shown through actions and behaviors.

-----

🛠️ Solution in this Paper:

→ LIIPA framework offers three variants for character analysis: LIIPA-DIRECT prompts LLMs directly for portrayal inference, while LIIPA-STORY and LIIPA-SENTENCE generate intermediate word lists that map to portrayal labels.

→ The framework analyzes characters across three dimensions: intellect, appearance, and power, classifying each as low, neutral, or high.

→ ImPortPrompts dataset was created with greater cross-topic similarity and lexical diversity than existing narrative datasets.

-----

💡 Key Insights:

→ LLMs can effectively analyze implicit character portrayal in complex narratives

→ There exists a fairness-accuracy tradeoff in LLM-based character analysis

→ Full narrative context improves portrayal analysis accuracy

-----

📊 Results:

→ LIIPA outperforms existing COMET-based approaches in both fairness and accuracy

→ LIIPA maintains performance with increasing character counts due to full context utilization

→ All LIIPA variants show better fairness-accuracy balance than non-LLM baselines