Plug-and-Blend: A Framework for Controllable Story Generation with Blended Control Codes
This work addresses the challenge of controlling language models for creative applications like story generation, offering a method for users to manage topics and transitions, though it appears incremental as it builds on existing controllable generation techniques.
The authors tackled the problem of controlling large language models for creative story generation by introducing Plug-and-Blend, a framework that allows users to input multiple control codes for topics, enabling loose or fine-grained control over topic transitions and blending. Automated evaluations demonstrated that the framework effectively guides generation towards given control codes while maintaining fluency, and human evaluations showed observable topic transitions in generated stories.
Large pre-trained neural language models (LM) have very powerful text generation capabilities. However, in practice, they are hard to control for creative purposes. We describe a Plug-and-Play controllable language generation framework, Plug-and-Blend, that allows a human user to input multiple control codes (topics). In the context of automated story generation, this allows a human user loose or fine-grained control of the topics and transitions between them that will appear in the generated story, and can even allow for overlapping, blended topics. Automated evaluations show our framework, working with different generative LMs, controls the generation towards given continuous-weighted control codes while keeping the generated sentences fluent, demonstrating strong blending capability. A human participant evaluation shows that the generated stories are observably transitioning between two topics.