基于强化学习的古今汉语句子对齐研究

Summary

Neural machine translation between Classical (ancient) and Modern Chinese lacks large, high-quality sentence-aligned parallel data. Older alignment methods score sentence pairs with length, lexical, and shared-character features. They capture meaning poorly and struggle with many-to-many alignments. The authors frame alignment within already-matched paragraph pairs as an unsupervised sequential decision problem.

A BERT model works as a binary matcher, using a next-sentence-prediction-style input of the form [CLS] ancient [SEP] modern. It is first warmed up on rough sentence pairs that the LCS algorithm extracts, with randomly sampled negatives added. It is then trained as a policy network inside a dynamic-programming search over 1-1, 1-2, 2-1 and 2-2 candidate merges. The training signal is a reward for covering as many sentence pairs as possible, optimized with policy gradient plus a baseline. Matches involving more than one sentence earn larger rewards, which pushes the model to notice non-1-1 patterns.

On a manually aligned test set from the Twenty-Four Histories, the approach beats the strongest earlier baseline. The biggest gains are on many-to-many alignments. This provides a way to mine parallel data for low-resource historical translation.

Key points

• Data: About 300 articles from 《二十四史》 (Twenty-Four Histories) were crawled with their modern translations, giving roughly 9,000 hand-aligned paragraph pairs.
  • 200 of these paragraphs were sentence-aligned by hand, yielding 1,550 test sentence pairs labeled by alignment pattern.
  • The remaining paragraphs serve as training data.
• Alignment patterns: Earlier statistics show 1-1 alignments make up about 90.12% of the data, and patterns up to 2-2 cover about 98.98%. The model therefore handles only patterns up to 2-2, and folds 1-0/0-1 cases into 1-2/2-1.
• Rewards and complexity: A rejected match earns 0 reward, an accepted 1-1 match earns 2, and accepted 1-2, 2-1 and 2-2 matches earn 3 each. The DP runs in O(MN) time, so the method suits paragraph-aligned corpora rather than whole-document ones.
• Claimed contributions:
  • first use of a pre-trained language model for ancient–modern Chinese sentence alignment;
  • a global RL objective that maximizes alignment coverage;
  • a newly built alignment corpus.
• Publication: The PDF header shows the paper appeared in 中文信息学报 vol. 38, no. 12 (Dec 2024), pp. 30–38. It comes from Beijing Language and Culture University.