Autoregressive Poetry Generation with Attention-Based Rule Adherence

The patent transforms discrete poetic rules (rhymes, tonal patterns, antithesis forms) into continuous parameter spaces that can be optimized by neural networks. The language model learns to generate poetry by adjusting parameters such as character probabilities, word sequences, and structural constraints, enabling consistent adherence to poetic rules while maintaining artistic quality across multiple generated poems.

The patent replaces traditional mechanical approaches to poetry generation (rule-based systems, template matching) with a neural network-based language model. This substitution enables the system to learn complex poetic patterns from training data and generate consistent, high-quality poetry that adheres to traditional rules without requiring explicit programming of each rule.

The patent segments the poetry generation process into multiple sequential steps performed by different components of the language model: token prediction, sequence generation, and poetic rule validation. This segmentation allows each component to specialize in specific aspects of poetry generation, maintaining high coherence while enabling efficient parallel processing and fast generation speeds.

The patent implements continuous refinement of generated poetry through iterative prediction and validation cycles. The language model continuously generates and refines poem sequences, maintaining coherence throughout the generation process by constantly evaluating and adjusting predictions against learned poetic patterns, thereby achieving both high quality and efficient generation.