Compose Like Humans: Jointly Improving the Coherence and Novelty for Modern Chinese Poetry Generation

Lei Shen, Xiaoyu Guo, Meng Chen

Abstract

Chinese poetry is an important part of worldwide culture, and classical and modern sub-branches are quite different. The former is a unique genre and has strict constraints, while the latter is very flexible in length, optional to have rhymes, and similar to modern poetry in other languages. Thus, it requires more to control the coherence and improve the novelty. In this paper, we propose a generate-retrieve-then-refine paradigm to jointly improve the coherence and novelty. In the first stage, a draft is generated given keywords (i.e., topics) only. The second stage produces a "refining vector" from retrieval lines. At last, we take into consideration both the draft and the "refining vector" to generate a new poem. The draft provides future sentence-level information for a line to be generated. Meanwhile, the "refining vector" points out the direction of refinement based on impressive words detection mechanism which can learn good patterns from references and then create new ones via insertion operation. Experimental results on a collected large-scale modern Chinese poetry dataset show that our proposed approach can not only generate more coherent poems, but also improve the diversity and novelty.

Introduction

While AI has become quite adept at generating classical Chinese poetry—which relies on strict rules for length, rhyme, and tonal patterns—modern Chinese poetry presents a unique set of challenges. Modern poetry is flexible, free-form, and lacks rigid constraints, making it difficult for models to maintain a coherent theme across multiple lines. Furthermore, existing models often suffer from "boring" outputs, either generating generic sentences or failing to introduce evocative imagery. Previous "refinement" approaches usually only considered immediate local context, failing to account for the global thematic direction of the poem.

The GRR Framework

To address these issues, we propose the GRR (Generate-Retrieve-then-Refine) paradigm, which mimics the human creative process by separating ideation from polishing. The framework consists of three distinct stages:

1. Draft Generation

Given a set of keywords representing the poem's topics, the model first generates an initial "rough draft." This draft serves as a global roadmap, providing future sentence-level information for the lines to be generated.

2. Impressive Words Detection

The core innovation lies in distilling "impressive words" from human-written masterpieces. The model uses the draft and keywords to search a massive database of human poetry. It identifies "impressive" nouns, adjectives, and verbs that are present in high-quality human poems but missing from the initial AI draft. These words are encoded into a "refining vector."

3. Refinement

The final poem is generated by a decoder that balances three critical inputs: bidirectional context (looking at both past and future lines from the draft), the draft line itself, and the refining vector to inject sophisticated vocabulary.

The architecture of the Refinement stage in the GRR framework, illustrating how bidirectional context and impressive words are integrated into the final decoder.

Dataset and Implementation

Because no large-scale public dataset for modern Chinese poetry existed, the authors built a new corpus comprising 263,669 poems (over 9.2 million sentences). This dataset includes specialized modern poetry and song lyrics, which share similar emotional depth and stylistic flexibility. Keywords for training were automatically extracted using the TextRank algorithm to facilitate the draft generation process.

Experimental Results

The GRR model was tested against several strong baselines, including Planning-based models and Deliberation Networks.

Quantitative and Qualitative Performance

As shown in the evaluation results, GRR achieved the highest Rouge-L (0.4138) and significantly better Novelty scores compared to baselines. In human evaluations conducted by literature postgraduates, GRR consistently outperformed other models in Coherence (3.98/5), Impressiveness (3.86/5), and Poeticness (3.40/5).

Automatic and human evaluation results showing that GRR outperforms baseline models across nearly all metrics, including fluency, coherence, and novelty.

Key insights from the results include:

  • Sentence-Level Lookahead: By utilizing the draft, the model can "see the future," allowing for tighter thematic consistency than standard one-pass models.
  • Distillation vs. Retrieval: Instead of noisy "copy-pasting," GRR distills the essence of retrieved lines, allowing the generative model to weave them naturally into the text.

Conclusion

The GRR framework represents a significant step toward human-like creativity in AI poetry generation. By separating the drafting and refinement phases, the model achieves superior coherence and novelty. While a gap remains between AI and human poets—particularly in deep metaphorical resonance—the "draft-and-refine" logic provides a robust foundation for future work in creative text generation.

Summary updated

Abstract

Chinese poetry is an important part of worldwide culture, and classical and modern sub-branches are quite different. The former is a unique genre and has strict constraints, while the latter is very flexible in length, optional to have rhymes, and similar to modern poetry in other languages. Thus, it requires more to control the coherence and improve the novelty. In this paper, we propose a generate-retrieve-then-refine paradigm to jointly improve the coherence and novelty. In the first stage, a draft is generated given keywords (i.e., topics) only. The second stage produces a "refining vector" from retrieval lines. At last, we take into consideration both the draft and the "refining vector" to generate a new poem. The draft provides future sentence-level information for a line to be generated. Meanwhile, the "refining vector" points out the direction of refinement based on impressive words detection mechanism which can learn good patterns from references and then create new ones via insertion operation. Experimental results on a collected large-scale modern Chinese poetry dataset show that our proposed approach can not only generate more coherent poems, but also improve the diversity and novelty.
BibTeX
@misc{shen2020compose,
  author = {Lei Shen and Xiaoyu Guo and Meng Chen},
  title = {{Compose Like Humans: Jointly Improving the Coherence and Novelty for Modern Chinese Poetry Generation}},
  year = {2020},
  eprint = {2005.01556},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.01556},
}
Figure