Elamur Research⌕
☰
home / papers / 2026 / Beyond Prompt-Sensitive Emotion Words: Stable Embeddings for Tang Poetry Analysis

Beyond Prompt-Sensitive Emotion Words: Stable Embeddings for Tang Poetry Analysis

Paper · 2026 · Linyue Zhang and Feiyue Li

Summary

Emotion research on Tang poetry has mostly relied on broad polarity categories. Newer LLM methods that ask a model for a single emotion word are fragile, because small changes to the prompt alter the answer, so historical conclusions built on those labels are hard to reproduce. Zhang and Li address this for poems from around the An Lushan Rebellion (754–764). Instead of collecting generated words, they treat each line's last-layer hidden state as a sentence embedding and cluster these vectors with K-means. They then merge the clusters by hand into a fine-grained multi-label scheme, which is further mapped onto seven macro-emotions to make trends easier to read. Two prompt variants assign the same word to only about half of the lines, while the embedding clusters stay evenly populated. Simple baselines show that the label set is learnable. Year- and poet-level aggregation identifies 762 as the main emotional turning point. The authors argue that stable representations are a precondition for treating computational output as credible humanities evidence.

Key points

  • Corpus: 616 poems by 47 poets (9,965 sentences, expanded to 10,122 rows after manual line splitting), of which 7,195 rows are labeled. Annotator agreement on a 200-sentence sample reached OA 0.970 and Cohen's κ 0.796.
  • Label construction: the elbow criterion selected k = 20 clusters, which were manually consolidated into 18 labels (0–3 per sentence). These were then mapped to 7 macro-categories based on the Dalian University of Technology emotion ontology lexicon.
  • Prompt sensitivity: on 3,198 emotional sentences, two prompt templates with DeepSeek-V3.2 at T=0 produced only 50.3% exact agreement. Embedding clustering achieved normalized entropy of 0.989 with all 20 clusters active.
  • Supervised baseline: a character-based TF-IDF model reached 0.446 micro-F1 and 0.395 macro-F1 on the 18-label task.
  • Historical findings:
    • In 762, Sorrow (0.7024) and Melancholy (0.1824) peak, while Romantic Affection drops to its minimum (0.0049).
    • From 757 onward, Heroic Aspiration declines.
    • Du Fu leans toward grief, whereas poets such as Dugu Ji and Liu Changqing show relatively high Longing.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSocial Sciences · 1,954 worksGeneral Social Sciences · 254 worksLiterature and Literary Theory · 541 works

🧭 Topics 1Research subjects and languages the work is about.Sentiment & emotion · 1,035 works

Cite

Cite as Linyue Zhang and Feiyue Li. 2026. Beyond Prompt-Sensitive Emotion Words: Stable Embeddings for Tang Poetry Analysis. Proceedings of the 6th International Conference on Natural Language Processing for the Digital Humanities.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.nlp4dh-1.7

@inproceedings{zhang2026beyond,
  title = {{Beyond Prompt-Sensitive Emotion Words: Stable Embeddings for Tang Poetry Analysis}},
  author = {Linyue Zhang and Feiyue Li},
  booktitle = {Proceedings of the 6th International Conference on Natural Language Processing for the Digital Humanities},
  year = {2026},
  doi = {10.18653/v1/2026.nlp4dh-1.7},
  url = {https://doi.org/10.18653/v1/2026.nlp4dh-1.7}
}

References

7 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. David Bamman, Ted Underwood, and Noah A. Smith. 2014. A Bayesian mixed effects model of literary character. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 370–379, Baltimore, Maryland. Association for Computational Linguistic…
  2. Sven Buechel and Udo Hahn. 2017. EmoBank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 578–585, Va… find
  3. Renfen Hu and Yuchen Zhu. 2015. Automatic topic classification for tang poems. Journal of Peking University (Natural Science Edition), 51(2):262–268. [in Chinese]. find
  4. Institute of Information Retrieval, Dalian University of Technology. 2007. Dalian university of technology emotion ontology lexicon. urlhttps://ir.dlut.edu.cn/zyxz/qgbtk.htm. [in Chinese]. find
  5. Guang Sima. 1956. Zizhi Tongjian. Zhonghua Book Company, Beijing. [in Chinese]. find
  6. Xue Xiao. 2006. Hierarchical classification of chinese text and its application to tang poetry classification. Master’s thesis, Chongqing University. [in Chinese]. find
  7. Wei Zhang, Hao Wang, Sanhong Deng, and Baolong Zhang. 2021. Emotion-term extraction and application for classical chinese poetry in digital humanities. Journal of Library Science in China, 47(4):113–131. [in Chinese]. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-beyond-prompt-sensitive-emotion-words/