2026-10-15T03:00:08Z https://www.scidb.cn/oai

10.57760/sciencedb.32366 2026-09-20T11:20:24Z

2026-09-20 A dataset of Tang poetry sentence pairs for metrical imitation doi:10.57760/sciencedb.32366 en Large Language Models (LLMs) often exhibit deficiencies in tonal patterns (pingze), structural parallelism, and imagery diversity when performing stylistic imitation of Tang poetry. To address these challenges, this study presents a specialized paired corpus of Tang poetry imitation used for metrical imitation tasks. Sourced from The Complete Tang Poems (Quan Tangshi), the dataset comprises 17,960 entries across four distinct genres: pentasyllabic/heptasyllabic quatrains, pentasyllabic regulated verses, heptasyllabic quatrains, and heptasyllabic regulated verses. Each entry consists of an original poetic couplet line pair and a corresponding manually crafted imitation line pair. All imitation line pairs were created by specialists in Classical Chinese. While strictly following the metrical patterns of the original lines, the generated imitations underwent a triple-verification process—covering metrical conformity, antithetical parallelism, and imagery consistency, ensuring that no core content words were repeated from the original lines. A dual-person review of a randomly selected 5% sample showed that 94.9% of the samples passed all three evaluation criteria: metrical conformity, antithetical parallelism, and imagery consistency. For the Qwen3-8B, metrical accuracy increased from 32.8% to 62.8%, the pass rate improved from 66.8% to 86.2%, and the imagery diversity score improved from 0.48 to 0.76. This dataset can be used for supervised training of regulated poetry generation models and offers data resources for text generation tasks involving strong formal constraints. Tang poetry; metrical imitation; human-authored imitation; dataset Li Ailin Wang Jizong Shangguan Wentong Wang Zhihang Hu Rong Peng Bo Sheng Wei Wang Wenjun PUBLIC https://creativecommons.org/licenses/by-nc/4.0/ dataset http://www.doi.org/10.11922/11-6035.csd.2025.0219.zh Science Data Bank