Go to ACL ARR 2026 January homepageMASES: A Multi-Agent Framework for Expert-Level Evaluation of Film Script
Keywords: Multiagent Systems under Uncertainty
Abstract: In the process of film production, script evaluation is a crucial but costly stage, traditionally relying on expert judgment from the subjective narrative dimension. Meanwhile, the rapid growth of automatically generated scripts has increased the demand for scalability and systematic methods for evaluating script quality. Based on the classical script theory, we propose a multi-agent framework for expert-level film script evaluation. MASES breaks down script evaluation into six core narrative dimensions, each of which is evaluated by a dedicated agent to generate structured reviews. We have also released a new dataset that contains over 100 feature film-length scripts, annotated by professional teachers and supplemented by audience comments consistent with the dimensions. We further validate the causal reasoning of MASES by aligning its generated narrative causal chains with expert and crowd annotations. Experiments show that MASES is well consistent with the evaluations of both experts and audiences, supporting scalable and interpretable script evaluations.
Paper Type: Long
Research Area: AI/LLM Agents
Research Area Keywords: LLM agent
Contribution Types: Data resources, Data analysis
Languages Studied: English
Submission Number: 8884
Loading