Creative work used to be human-exclusive. That’s no longer the case, as AI has demonstrated its ability to create paintings, compose songs, and write news stories. In a new expression of its ever-widening artistic ability, AI is generating and refining traditional Chinese poems.
Researchers from the University of Science and Technology of China and Ping An Technology recently proposed a novel framework for high quality Chinese poetry generation in the paper An Iterative Polishing Framework based on Quality Aware Masked Language Model for Chinese Poetry Generation.
Beloved across time and space, classical Chinese poetry dates back 30 centuries. The great poems — romantic or realistic — are developed from only a few hanzi characters and some fixed rules and forms.
The importance of character choice is highlighted in a story of revered Tang Dynasty poet Jia Dao, who while writing On Li Ning’s Secluded Residence, struggled so hard between tui (pushes the door) and qiao (knocks at the door) for the phrase “Birds nestle in the trees by the pond, a monk knocks at the door awash in the moon,” that “tuiqiao” became a Chinese phrase meaning to act with repeated deliberation.
Due to the nuanced character choices and other unique literal and aesthetical characteristics, automatic generation of Chinese poetry is challenging for AI, and high-quality poems can hardly be generated by end-to-end methods. The researchers therefore proposed a framework consisting of a BERT-based encoder-decoder structure to first generates a poem draft, and a new Quality-Aware Masked Language Model (QA-MLM) to polish the draft towards higher quality in terms of linguistics and literalness.
QA-MLM was also designed to evaluate a poem’s quality in terms of semantics, syntactics and literary. It’s not only able to locate improper characters but also to replace them with better ones by synthesizing the all-round poem context information. QA-MLM will automatically terminate the polishing process when the edited draft is classified as qualified.
Trained and tested with about 130,000 poems from five dynasties, three models (key2one, one2one, and two2one) were used to generate the poem drafts. Then QA-MLM came in to perform further polishing.
Both evaluation by humans and automatic evaluation metrics including BLEU, tone accuracy, and rhythm accuracy showed that the new approach is effective in Chinese poetry generation. Moreover, their proposed QA-MLM can even improve the consistency, fluency, meaningfulness and poeticness of poems generated by the other three models tested.
Although QA-MLM research is currently focused only on Chinese poetry generation, its text refinement approach could be extended to other natural language generation areas. For writers of today whose struggles with word choice echo Jia Dao’s through the ages, the algorithmic editing input could be welcome.
The paper An Iterative Polishing Framework based on Quality Aware Masked Language Model for Chinese Poetry Generation has been accepted by AAAI 2020 and is available on arXiv.
Journalist: Yuan Yuan | Editor: Michael Sarazen