Elamur Research⌕
☰
home / papers / 2026 / YNU-HPCC at SemEval-2026 Task 4: Narrative Similarity via Multi-Perspective E5-Mistral and Embedding Routing

YNU-HPCC at SemEval-2026 Task 4: Narrative Similarity via Multi-Perspective E5-Mistral and Embedding Routing

Paper · 2026 · Feiyang Song, Jin Wang and Xuejie Zhang

Summary

The authors tackle SemEval-2026 Task 4, which asks whether two stories share the same underlying narrative (abstract theme, course of action, outcomes) regardless of wording, characters or setting. Standard dense encoders tend to reward lexical overlap, so they struggle with pairs that look alike on the surface but resolve differently.

For Track A (triplet ranking), the team fine-tuned E5-Mistral-7B-Instruct with LoRA and MNRL on 10,000 synthetic triplets generated by Gemini-2.5-Pro. Each story is encoded as three fused views: theme, verb-based action, and ending-based outcome. When the cosine-similarity gap between the two candidates is small, the triplet is passed to Gemini-3-Flash, which reasons with chain-of-thought and majority voting.

For Track B (representation learning), they used Gemini-Embedding-001 on head-body-tail excerpts and then applied ABTT post-processing.

The system placed 5th in Track A with 0.7425 accuracy and 2nd in Track B with 0.7125. The results suggest that routing uncertain cases from a cheap encoder to an LLM, combined with large general-purpose embedders, works well for narrative-level similarity.

Key points

  • Track A routing: the view embeddings are weighted 0.5 (theme), 0.25 (action) and 0.25 (outcome). Triplets with a margin below τ = 0.072, a value tuned by grid search on the dev set, go to Gemini-3-Flash, sampled three times at temperature 1.0 with majority voting.
  • Dev-set scores:
    • E5-Mistral rose from 0.670 zero-shot to 0.705 after fine-tuning.
    • On the hard subset, Gemini-3-Flash (0.730) beat GPT-5.2 (0.695) and Qwen3-Max (0.675).
    • The full hybrid reached 0.750 on the dev set.
  • Data augmentation: ROCStories and the CMU Book Summary dataset were dropped because of domain shift. The final training data was LLM-synthesized hard negatives that follow the official annotation guidelines.
  • Track B setup: Gemini-Embedding-001 (768 dimensions, task type SEMANTIC_SIMILARITY) encodes the first 3, 3 sampled middle, and last 3 sentences of each story. ABTT then removes the first principal component from the embeddings. This setup outperformed the fine-tuned E5-Mistral, which scored 0.6600 on the test set.
  • Official baselines for comparison:
    • Track A: GPT-4o-mini scored 0.6700.
    • Track B: story-emb scored 0.6325.
    • Future work: distilling Gemini-3-Flash's reasoning into small local bi-encoders.

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

 Institutions

🎓Yunnan UniversityKunming · China17 works · 40 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 worksInformation Systems · 1,437 works

Cite

Cite as Feiyang Song, Jin Wang and Xuejie Zhang. 2026. YNU-HPCC at SemEval-2026 Task 4: Narrative Similarity via Multi-Perspective E5-Mistral and Embedding Routing. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.semeval-1.63

@inproceedings{song2026ynu,
  title = {{YNU-HPCC at SemEval-2026 Task 4: Narrative Similarity via Multi-Perspective E5-Mistral and Embedding Routing}},
  author = {Feiyang Song, Jin Wang and Xuejie Zhang},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.63},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.63}
}

References

17 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. David Bamman and Noah A. Smith. 2013. New alignment methods for discriminative book summarization. CoRR, abs/1305.1319. — [arXiv:1305.1319](https://arxiv.org/abs/1305.1319) find
  2. Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of the Association for Computational Linguistics: ACL 2024, pages 2318–2335, Bang…
  3. Hans Ole Hatzel, Ekaterina Artemova, Haimo Stiemer, Natalia Fedorova, Evelyn Gius, and Chris Biemann. 2026. Narrative similarity-annotation guidelines. Technical report. find
  4. Matthew L. Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, László Lukács, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. 2017. Efficient natural language response suggestion for smart reply. CoRR, abs/1705.00652. find
  5. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In ICLR. OpenReview.net.
  6. Weijie Li, Jin Wang, and Xuejie Zhang. 2024. YNU-HPCC at SemEval-2024 task 1: Self-instruction learning with black-box optimization for semantic textual relatedness. In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), pages 792–799, Mexico City, Mexico. Associati… find
  7. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. find
  8. Yuheng Mao, Jin Wang, and Xuejie Zhang. 2025. YNU-HPCC at SemEval-2025 task 7: Multilingual and cross-lingual fact-checked claim retrieval. In Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2025), pages 234–240, Vienna, Austria. Association for Computational Linguisti… find
  9. Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 find
  10. Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. MTEB: Massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2014–2037, Dubrovnik, Croatia. Association for Computational Linguistic…
  11. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. CoRR, abs/1908.10084.
  12. Qwen Team. 2025. Qwen3 technical report. Technical report. find
  13. Hannah Waight, Sol Messing, Anton Shirikov, Margaret E Roberts, Jonathan Nagler, Jason Greenfield, Megan A Brown, Kevin Aslett, and Joshua A Tucker. 2025. Quantifying narrative similarity across languages. Technical report. find
  14. Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. CoRR, abs/2212.03533.
  15. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11897–11916, Bangkok, Thailand. Asso…
  16. Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In SIGIR 2024 - Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 641–649. Assoc…
  17. Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, and Min Zhang. 2024. mGTE: Generalized long-context text representation and reranking models for multilingual text retrieval. In Proceedings of the 2024…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/song-2026-ynu-hpcc-semeval-2026-task/