A residual GRU-Attention hybrid architecture for high-fidelity Chinese classical poetry generation

Department of Computer Science, University of Utah, Salt Lake City, UT, USA

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Automatic generation of classical poetry is a highly challenging constrained text generation task, demanding strict adherence to stylistic rules like meter, rhyme, and parallelism alongside high-level semantic coherence. While hybrid architectures combining Recurrent Neural Networks (RNNs) and Attention mechanisms show promise, they often suffer from significant training instability, hindering their performance. This paper systematically investigates these stability challenges in RNN-Attention hybrid architectures and demonstrates through comparative experiments the decisive impact of architectural integration strategies on model convergence and generation quality. We validate our findings on Chinese classical poetry generation, a domain characterized by its intricate constraints. Experimental results show that our proposed model (V3), which employs a novel integration strategy, achieves a comprehensive quality score of 8.1, marking a significant improvement of 24.6% over a strong GRU baseline (V2). Furthermore, ablation studies confirm that a naive concatenation of modules leads to mode collapse, highlighting that the proposed integration strategy— rather than the simple stacking of components—is the key to ensuring model stability and achieving superior performance. Our work offers valuable insights for designing robust hybrid models for other complex text generation tasks.