Collaborative and Controllable AI for Creative Co-Creation

Creative work often thrives under structure: melodies guide lyrics, rhyme constrains expression, and collaboration enhances imagination. Yet most generative models today remain either fully autonomous or narrowly optimized, lacking the ability to co-create with humans under meaningful constraints. This talk explores how we can design collaborative and controllable AI systems that augment rather than replace human creativity. Through the lens of lyric generation, I will present three complementary directions. Unsupervised Melody-to-Lyric Generation introduces a hierarchical framework that composes lyrics guided by musical structure without any parallel training data, enabling controllable creativity under data scarcity. REFFLY extends this paradigm into a revision-based model that edits plain-text drafts into melody-aligned lyrics while preserving both semantics and musicality. Finally, CoLyricist bridges modeling and design, offering a workflow-aligned songwriting assistant that supports human lyricists across ideation, drafting, and melody fitting. Together, these works illustrate a broader vision for AI as a creative partner: systems that learn to respect structure, understand human intention, and engage in genuine co-creation.