基于词性对齐与依存关系的中文排比句生成方法

Summary

Parallelism is a rhetorical device in which three or more structurally matching clauses appear side by side. Chinese parallelism generation had no dedicated models and no public corpus. The authors, from Jiangxi Normal University, address this task: given one clause, produce follow-up clauses that match it in length, punctuation, syntax and meaning.

They first compiled a corpus of parallel sentences. They then built CPG-PosDep, a Transformer-based autoregressive generator with two main parts:

• Part-of-speech alignment component. Every sentence group receives three kinds of global tags: word segmentation, word order within a clause, and clause position. The model learns these tags through masked and global multi-head attention.
• Dependency component. BERT and an attention module encode the dependency structure of the given clause.

The two feature streams are fused to generate the new clauses. On BLEU and Distinct, the generated clauses kept corresponding-position POS and dependency relations consistent with the input. Their quality also exceeded output from existing couplet and poetry generators. This matters because those generators focus on rhyme and tone rather than structural parallelism, so they transfer poorly to this task.

Key points

• New resource: "Parallelisms Datasets," with 30,452 groups of Chinese parallel sentences, is released on GitHub (puppy210/Parallelisms-Datasets).
• Preprocessing: Stanford CoreNLP segments and POS-tags the corpus, which uses 31 POS categories (e.g., VV, VE, NN, PU).
• Three tag types: The tags are summed into a single feature vector F.
  • Segmentation tags (C) mark word boundaries and word lengths.
  • Reverse-indexed in-clause position tags (P) help the model detect where a clause ends.
  • Ascending clause-index tags (S) distinguish one clause from another.
• Random sampling strategy: One word in the given clause is chosen at random, and the aligned slots in later clauses show the actual word instead of a placeholder. This teaches the model both repeated words and same-POS substitutions.
• Evaluation: The model is compared with couplet and poetry generation baselines using BLEU and Distinct. The paper reports clear quality gains, but the extracted text omits the specific scores.