Introduction
Ancient Chinese Ci poetry represents a pinnacle of literary art, defined by its evocative imagery and rigid structural constraints. Unlike many poetic forms with uniform line lengths, Ci follows specific templates known as Cipai. Each Cipai mandates a precise rhythmic structure, tonal patterns (alternating between "Ping" or level tones and "Ze" or slant tones), and strict rhyming schemes.
While modern neural networks have shown promise in creative text generation, they often struggle with these "hard" constraints. Standard models typically attempt to learn these rules implicitly from data, frequently resulting in poems that may sound poetic but fail the rigorous formal tests of classical composition. Researchers have addressed this gap with Metrically Restricted Ci Generation (MRCG), a model designed to explicitly obey the mathematical rigor of Ci while capturing its semantic beauty.
Methodology: The MRCG Framework
The MRCG model is built upon a Conditional Variational Auto-Encoder (CVAE) architecture. The core innovation lies in how the model handles the metrical constraints. Instead of hoping the decoder learns the rules through exposure, the researchers explicitly feed the metrical structure into the generation process as a condition.
Encoding and Latent Space
The encoder processes existing poems through a hierarchical structure: character-level embeddings are aggregated into line-level representations via a Gated Recurrent Unit (GRU), which are then synthesized into a context embedding. This process yields a latent semantic representation ($z$), capturing the thematic essence of the poem.
Explicit Metrical Conditioning
The "Condition" ($r$) is the defining feature of MRCG. The metrical rules of a chosen Cipai are converted into vectors:
- Tone-rule vectors: Specify whether a character at a given position must be "Ping," "Ze," or is flexible.
- Rhyme-rule vectors: Identify "rhyming feet" and their corresponding rhyme groups.
The decoder is then conditioned on both the semantic latent vector ($z$) and these metrical vectors ($r$). By integrating these rules directly into the decoding step, the model is forced to select characters that satisfy the structural requirements of the Cipai without sacrificing linguistic fluency.
Experimental Results
The model was trained on a dataset of over 82,000 Ci poems spanning 818 different Cipai. To evaluate performance, the researchers compared MRCG against standard Sequence-to-Sequence (Seq2Seq) and Attention-based models.
The results demonstrated a significant breakthrough in structural fidelity. While baseline models failed to generate poems with correct lengths or tones more than half the time, MRCG achieved near-perfect accuracy across all formal metrics.
Beyond objective metrics, human evaluation by experts confirmed that the enforcement of strict rules did not degrade the aesthetic quality of the poetry. MRCG-generated poems were rated higher in overall quality (2.45/4.0) compared to the Attention baseline (2.21), suggesting that the structural constraints actually help the model produce more authentic-sounding classical verse.
Conclusion and Future Directions
The MRCG model represents a successful fusion of "symbolist" AI (rule-based logic) and "connectionist" AI (neural learning). By treating poetic rules as explicit conditions rather than implicit patterns, the model bridges the gap between formal correctness and creative expression.
Future work aims to address remaining challenges, such as "implicit phrasing" (sub-rhythms within lines) and improving long-range narrative coherence. As it stands, MRCG provides a robust framework for preserving and generating one of the world's most complex literary forms through modern machine learning.
Summary updated