SNAP: Enhancing Long-Form Narrative Agents with Cell-Based Segmentation and Plan-Driven Dialogue Strategies
Abstract: Large Language Models (LLMs) hold great potential across domains like gaming, education, and cultural content. However, ensuring character consistency and narrative coherence in extended storytelling remains challenging. We propose SNAP (Story and Narrative based Agent with Planning), a framework that segments narratives into Cells, defines explicit Plans for each Cell, and guides dialogue generation accordingly. By limiting context within each Cell and providing plans with clear spatiotemporal settings, character actions, and plot developments, SNAP enables consistent, goal-driven dialogues. Experiments with novel datasets and human evaluations show that SNAP outperforms Vanilla GPT-4o-based agents in linearity, continuity, appropriateness, and non-redundancy, demonstrating its effectiveness in creating immersive long-form conversational agents.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: Dialogue and Interactive Systems, Generation, Human-Centered NLP, Language Modeling, NLP Applications
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 878
Loading