Co-DIRECT: A Multi-agent based Observable Drama Script Generation and Performance

ACL ARR 2025 February Submission1633 Authors

Abstract: Multi-agents cooperation framework shows excellent ability in storytelling and role-playing, but current end-to-end methods often generate homogeneous drama scripts by focusing solely on plot coherence, neglecting the diverse interpretive and imitation capabilities of multiple agents. This paper introduces Co-DIRECT, an enhanced Director-Writer-Actor-Critic multi-agent framework designed for human-AI co-creation in drama script generation and performance. Human directors provide story settings and observe the creative process, while specialized agents - Writer, Actor, and Critic - collaborate to generate, perform, and evaluate the script. Co-DIRECT injects a knowledge graph of classic scripts into the knowledge base, enabling agents to retrieve detailed information related to classic characters and archive the development context of classic plots holistically, thereby enhancing the agents' role-playing and plot generation capabilities. The use of digital human actors provides human directors with an observable interface. Experiments demonstrate a significant influence on human director engagement, underscoring the potential for collaborative creativity between humans and AI in storytelling.

Paper Type: Long

Research Area: Human-Centered NLP

Research Area Keywords: multi-agent, drama generation, retrieval-augmented generation, human-centered evaluation

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 1633