JiuGe: An Automatic Chinese Classical Poetry Generation System – Algorithms and Research Overview

The JiuGe system is a Chinese classical poetry automatic generation platform created by the THUNLP lab, aiming to explore machine intelligence and provide literary‑level poems that satisfy structural, semantic, and aesthetic requirements.

Motivation stems from the desire to study creative AI, following Turing's vision of machines writing poetry, and to investigate human intelligence through language, as poetry offers a rich testbed for computational creativity.

To improve literary quality, the system decomposes poetry generation into two layers: textual quality (coherence, relevance) and aesthetic features (novelty, style, emotion). Various algorithms address each layer, including a Salient Clue Model that selects semantically important tokens from the context, a Working‑Memory Model with history, local, and topic memory modules, and a Topic‑Memory with explicit keyword storage and tracking.

For style control, the authors introduce unsupervised learning with a regularization term that maximizes mutual information between style labels and generated text, enabling style‑conditioned generation despite limited labeled data.

Emotion control is achieved through semi‑supervised training that predicts emotion labels from topics and models both poem‑level and line‑level emotional dynamics using cross‑temporal sequence learning.

To overcome the limitations of maximum‑likelihood training, the system incorporates reinforcement learning with reward models that quantify human‑rated poetry metrics, and an inter‑generator reinforcement scheme that allows multiple generators to exchange information during training.

The complete pipeline, including knowledge‑graph integration (the “WenMi” graph) for entity linking and historical context, is deployed online (jiuge.thunlp.org), having generated over 25 million poems and receiving media coverage and awards.

Future work includes commonsense‑driven input understanding, deeper linguistic and literary knowledge incorporation, and spatio‑temporal historical grounding to further enhance poem generation.