BIPro: Zero-shot Chinese Poem Generation via Block Inverse Prompting Constrained Generation Framework (2411.13237v1)

Abstract: Recently, generative pre-trained models have made significant strides, particularly highlighted by the release of ChatGPT and GPT-4, which exhibit superior cross-domain capabilities. However, these models still face challenges on constrained writing tasks like poem generation under open-domain titles. In response to this challenge, we introduce Block Inverse Prompting (BIPro) constrained generation framework. BIPro leverages two block inverse prompting methods, revise and rewrite, that mimic the process of human text writing using block generative models. It significantly improves the zero-shot generation quality on the formidable constrained generation task of open-domain traditional-form Chinese poem generation. Based on a less powerful block generative model GLM-10B-Chinese, poems composed via BIPro without priming or additional training outperform both most advanced direct generative systems like GPT-4 or GLM-4 and best domain-specific systems such as Yusheng, Shisanbai, or Baidu Poetry Helper in human evaluation by proficient poets. Finally, BIPro considerably narrows the gap between AI-generated works and short-listed human literary arts in another human evaluation, unveiling the promising potential of block generative models in improving the quality of constrained generation.

• The paper introduces the BIPro framework, which uses block generative models with revise and rewrite methods for non-monotonic, constrained text generation.
• The paper demonstrates that BIPro significantly improves zero-shot Chinese poem quality, outperforming models like GPT-4 and GLM-4 in human evaluations.
• The paper highlights BIPro's broader implications for applications requiring rigorous stylistic adherence and iterative refinement in text generation.

An Overview of BIPro: A Constrained Generation Framework for Chinese Poetry

The study examined in this essay presents a novel approach to overcoming the challenges associated with generating traditional Chinese poetry using generative pre-trained models. The research introduces Block Inverse Prompting (BIPro), a constrained generation framework that addresses the inherent challenges posed by the rigid structures and aesthetic demands of traditional-form Chinese poetry.

The recent advancements in generative models like GPT-4 and ChatGPT have showcased their ability to perform across various domains, yet these models struggle with tasks involving strictly constrained text generation, such as poetry. This shortcoming arises mainly because conventional models generate text sequentially, focusing solely on preceding tokens and lacking mechanisms for revision once the text is produced. Poetry, especially in classical forms, necessitates a balance of constraints such as rhyme and meter, which require deliberate crafting and iteration.

The BIPro framework employs block generative models to enhance constrained generation. Specifically, this framework introduces two methods, "revise" and "rewrite," to simulate a human-like text creation process. The BIPro approach utilizes block generative models, which differ from traditional autoregressive models by allowing non-monotonic generation that considers both preceding and succeeding context during text production. This ability is crucial for poetry generation, where later lines can significantly influence the revision of earlier lines to achieve the intended artistic quality and structural coherence.

In empirical assessments, the BIPro framework applied to a relatively modest block generative model, the GLM-10B-Chinese, demonstrated notable improvement in poetic outputs compared to more powerful direct generative systems and specialized domain-specific systems. Utilizing zero-shot poem generation without additional training, the BIPro framework was able to outperform advanced systems like GPT-4, GLM-4, and others in human evaluations conducted by experienced poets. These results highlight the scaffold provided by BIPro in guiding generative models to produce text with complex constraints.

The implications of this research are significant for the development of AI systems capable of producing high-quality constrained text, such as poetry, which align closer to human artistic expression. It underscores the potential of integrating block generative models into broader applications, particularly those requiring adherence to strict literary or stylistic norms.

Looking toward future developments, enhancing the computational efficiency of the BIPro framework could make it more viable for large-scale applications. Furthermore, adapting this methodology to leverage more advanced base models could bridge the existing quality gap between AI-generated and human-generated poetry. Additionally, it offers exciting opportunities for exploration in other domains where constrained generation plays a pivotal role, such as legal document drafting, technical writing, or any field requiring stringent, stylistic adherence alongside creativity.

In summary, the BIPro framework emerges as a systematic advancement to improve the generation quality of AI models for constrained tasks such as Chinese poetry, leveraging the unique features of block generative models to simulate iterative, human-like refinement processes. The paper presents meaningful progress in AI's capacity for stylistically complex content creation, emphasizing the broader applicability of such constrained generation frameworks across diverse, real-world applications.