Generating Classical Chinese Poems from Vernacular Chinese

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Digital Humanities

Background:

• Classical Chinese poetry is a significant cultural heritage.
• Existing poem generation models offer limited user control over semantic meaning.
• Previous models primarily rely on keyword intervention.

Purpose of the Study:

• To propose a novel task for generating classical Chinese poems from vernacular text.
• To enhance user control over the semantic content of generated poems.
• To adapt unsupervised machine translation (UMT) techniques for this task.

Main Methods:

• Utilizing an unsupervised machine translation (UMT) framework.
• Implementing segmentation-based padding to mitigate under-translation.
• Employing reinforcement learning to address over-translation issues.

Main Results:

• Significant improvements in perplexity and BLEU scores compared to standard UMT models.
• Demonstrated effectiveness in generating high-quality classical Chinese poetry.
• Identified guidelines for optimizing input vernacular for superior poem generation.

Conclusions:

• The proposed method offers greater user control in classical Chinese poem generation.
• The approach achieves results comparable to amateur-written poems.
• This work advances the field of creative text generation in computational linguistics.