Most attempts to visualize what a poem “feels like” land on mood boards or static color palettes — a reasonable instinct, since poems don’t have a single emotion so much as a trajectory through several. A paper posted to arXiv on October 27, presented at the Student Research Consortium at India HCI 2026 (running October 29–31 at IIT Hyderabad, where this piece is landing as the conference opens), builds a pipeline that tries to represent that trajectory as motion instead of a snapshot.
The system, built by researcher Subhadra Nayak, works in two stages. First, a pretrained RoBERTa-based emotion classifier reads a poem and extracts its fine-grained emotional states — not just “sad” or “happy” broadly, but the finer gradations a classifier trained on emotional text can distinguish. Second, those detected emotional states get mapped onto strange attractors: the chaotic-but-structured mathematical systems (the Lorenz attractor being the best-known example) that generate the hypnotic, never-quite-repeating trajectories familiar from chaos-theory visualizations. Different emotional readings steer the attractor’s parameters, so the resulting visual shifts shape and behavior as the poem’s emotional content shifts.
The appeal of strange attractors specifically, rather than any other generative system, is that they’re already built to produce output that’s continuously varying without being random — a trajectory through an attractor never exactly repeats, but it’s also not noise. That property maps naturally onto something like a poem’s emotional arc, which moves and develops without simply jumping between discrete states.
The paper is six pages with five figures and carries no venue publication beyond the Student Research Consortium acceptance, so it’s worth reading as early-stage, conference-feedback-stage work — exactly what a consortium slot is for — rather than a finished system. Still, it’s a clean example of a pattern worth watching in generative art: using an NLP classifier not to generate text or images directly, but as a control signal feeding a separate, purpose-built visual system downstream.