📄 K-CoT：基于关键词思维链提示的中文排比句生成研究

Summary

Chinese parallelism (排比) consists of three or more clauses with matching syntax, a shared theme and a unified emotional tone. Research on generating it has been held back by two gaps: little high-quality data and no fine-grained annotation. LLMs also tend to produce clauses that are structurally misaligned or that drift off-topic. The authors address the data gap with CPDAK, a corpus drawn from prose, fiction, essays and student compositions. Each parallel sentence is annotated with its topic, parallel marker words, tone and keywords.

On top of this corpus they propose K-CoT, a prompting framework that needs no fine-tuning. It works in three steps:

• It selects diverse few-shot demonstrations through clustering.
• It derives keywords from the topic and tone through a step-by-step reasoning chain, with a semantic-similarity check.
• It produces the parallel sentence, then has the model reflect on its draft and revise it.

Across five open 6–9B LLMs, K-CoT beats plain prompting, standard CoT and a fine-tuned baseline. This points to keyword-mediated reasoning as an interpretable way to steer rhetorical generation.

Key points

• CPDAK corpus: about 10,023 sentences, 6,529 annotated parallelism instances and roughly 433K tokens. Each instance is labelled with a (Topic, Flag, Tone, Keywords) tuple. Inter-annotator agreement was measured with three annotators on 500 samples and reached Fleiss' Kappa = 0.78. The corpus is released on GitHub.
• Candidate filtering: candidates were pre-filtered by a Text2Vec-based RoBERTa Large classifier with a confidence threshold of 0.673. The classifier was trained on a balanced set of 5,195 examples, with couplets, repetition sentences and ordinary sentences serving as negatives.
• K-CoT demonstration selection: a dual-channel clustering combines BERT [CLS] embeddings with keyword embeddings (k=20 each, silhouette score > 0.6). Five demonstrations are drawn from it.
• K-CoT keyword filtering: generated keywords are kept only if their Sentence-BERT similarity to the topic exceeds 0.75.
• Models tested: ChatGLM3-6B, Baichuan2-7B, Qwen-2.5-7B, Yi-1.5-9B and LlaMA3-8B-Chinese-Chat.
• Baselines: CPGen-POS&DEP (fine-tuned), Vanilla prompting, and CoT with randomly chosen demonstrations.
• Evaluation: automatic metrics were BLEU and Distinct. Human raters scored five dimensions (structure, fluency, emotion, creativity, vividness) on a 1–3 scale.
• Headline result: with Qwen, K-CoT reached a BLEU-4 of 66.37, about a 107% gain over CPGen-POS&DEP, and improved the other metrics as well.