Abstract: Chinese parallelism research is challenged by two issues: the scarcity of high-quality corpora and the absence of fine-grained annotations. This study constructs a multi-dimensionally annotated Chinese Parallelism Corpus encompassing topic, tone, parallel markers, and keywords. Then, we propose K-CoT (Keyword-guided Chain-of-Thought), a novel generation framework that simulates human rhetorical composition through a progressive reasoning pipeline “topic deconstruction - feature mapping - keywords generation - syntactic synthesis”. Experimental results on mainstream models (such as ChatGLM, LlaMA) demonstrate that K-CoT achieves significant performance improvements in parallelism generation.