A Polishing Model for Machine-Generated Ancient Chinese Poetry

Zhe Chen,Yang Cao

DOI: https://doi.org/10.1007/s11063-024-11480-9

IF: 2.565

2024-03-03

Neural Processing Letters

Abstract:Machine poetry generation has been studied for decades, among which ancient Chinese poetry is still challenging in the field of poetry generation due to its unique regularity and rhythm. The quality improvement of ancient Chinese poetries is one of the most promising research areas of ancient Chinese Natural Language Processing. This paper proposes an ancient Chinese poetry polishing model, which is used for polishing to obtain high-quality ancient Chinese poetry. The model consists of a detection network and a correction network. The detection network based on BiLSTM and CRF is used to detect different types of low-quality words in poems. The correction network based on the BERT model is used to modify the detected low-quality words in the global context. The polishing process is iteratively performed until the model judges that there are no low-quality words in the poem. The results show that the polished poems are improved in multiple evaluations. Compared with existing polishing models, the model proposed in this paper performs better in both automatic evaluation and human evaluation when the number of parameters is reduced.

computer science, artificial intelligence

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve

This paper aims to address the issue of improving the quality of machine-generated ancient Chinese poetry. Specifically, due to its unique rhythm and regularity, ancient Chinese poetry remains challenging in the field of poetry generation. Most current poetry generation models rely on unidirectional generation using Recurrent Neural Networks (RNNs). Although this method can meet the formal requirements of poetry, there is still a significant gap in coherence and meaning expression. Additionally, existing poetry polishing methods often substantially alter the original content of the poem, leading to a shift in the original meaning.

To this end, this paper proposes an ancient Chinese poetry polishing model, which consists of a detection network and a correction network. The detection network is based on Bidirectional Long Short-Term Memory (BiLSTM) and Conditional Random Fields (CRF) and is used to detect different types of low-quality words in the poetry. The correction network is based on the BERT model and modifies the detected low-quality words by incorporating global context. The iterative polishing process continues until the model determines that there are no low-quality words left in the poetry. Experimental results show that the polished poetry significantly improves across multiple evaluation metrics, and the proposed model outperforms existing polishing methods in both automatic and manual evaluations, even with a reduced number of parameters.