Elamur Research⌕
☰
home / papers / 2026 / THIVLVC: Retrieval Augmented Dependency Parsing for Latin

THIVLVC: Retrieval Augmented Dependency Parsing for Latin

Paper · 2026 · Luc Pommeret, Thibault Wagret and Jules Deret

Summary

THIVLVC was built to take part in EvaLatin 2026's dependency parsing challenge. The task covers two genres: Seneca's poetry and Thomas Aquinas's philosophical prose. Supervised parsers tend to copy whatever annotation habits their training treebanks contain, so the authors instead let a large language model correct a conventional parser's output, guided explicitly by UD rules.

The pipeline works in two steps:

  • Retrieval: a structural retriever picks the five most similar sentences from the CIRCSE treebank, scoring candidates by sentence length and overlap of POS bigrams and trigrams.
  • Refinement: gemini-3-flash receives the UDPipe baseline parse, the retrieved examples and the UD guidelines, and outputs corrected HEAD and DEPREL columns in CoNLL-U.

On poetry the retrieval-augmented setup outperforms UDPipe by about 17 CLAS points, while on prose the gain is small.

A blind human review of 300 disagreements with the gold standard found that the system was preferred slightly more often than the gold. The reviewers also found inconsistencies within and across treebanks, which suggests that single-reference evaluation can penalize defensible analyses.

Key points

  • Configurations: THIVLVC_1 combines LLM, UDPipe and UD guidelines. THIVLVC_2 adds retrieval (RAG) over the CIRCSE training set (762 sentences, k = 5).
  • Retrieval quality: the structural retriever keeps the average length gap near one token, compared with more than 11 for TF-IDF and morphological retrieval. Its POS overlap remains competitive.
  • Poetry (CLAS, with subtypes): THIVLVC_2 scores 74.03, compared with 56.94 for UDPipe.
  • Prose (CLAS, with subtypes):
    • THIVLVC_2 scores 80.92, compared with 79.41 for UDPipe.
    • Without retrieval, THIVLVC_1 reaches only 74.04, so RAG adds about 6.9 points.
  • Error analysis: two annotators agreed unanimously on 167 of 300 items, and 53.3% of those favoured THIVLVC (Cohen's κ = 0.493). Recurring issues include:
    • advmod:lmod vs advmod conventions differing between CIRCSE and EvaLatin;
    • inconsistent use of obl vs obl:arg within the EvaLatin gold;
    • mark used where case is correct.

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

🔬Centre National de la Recherche ScientifiqueParis · France104 works · 60,877 citations🎓Université Paris-SaclayGif-sur-Yvette · France34 works · 5,733 citations🔬Laboratoire Interdisciplinaire des Sciences du NumériqueGif-sur-Yvette · France10 works · 422 citations🎓École Normale Supérieure de LyonLyon · France3 works · 236 citations

Also on the byline: Histoire et Sources des Mondes Antiques.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 9,085 worksArtificial Intelligence · 8,888 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksHistory · 66 works

🧭 Topics 3Research subjects and languages the work is about.Dependency parsing · 120 worksUniversal Dependencies · 174 worksLatin · 13 works

Cite

Cite as Luc Pommeret, Thibault Wagret and Jules Deret. 2026. THIVLVC: Retrieval Augmented Dependency Parsing for Latin. Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2q8twojtotyb

@inproceedings{pommeret2026thivlvc,
  title = {{THIVLVC: Retrieval Augmented Dependency Parsing for Latin}},
  author = {Luc Pommeret, Thibault Wagret and Jules Deret},
  booktitle = {Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/2q8twojtotyb},
  url = {https://doi.org/10.63317/2q8twojtotyb}
}

References

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

  1. Flavio Massimiliano Cecchini, Marco Passarotti, Paolo Ruffolo, Marinella Testori, Lia Draetta, Martina Fieromonte, Annarita Liano, Costanza Marini, and Giovanni Piantanida. 2018. Enhancing the Latin morphological analyser LEMLAT with a medieval Latin glossary. In Proceedings of the Fifth Italian Conference on Computational Linguistics (CLiC-it 2018), pages 88–93, Turin, Italy. CEUR Workshop Proceedings. find
  2. Marie-Catherine de Marneffe, Timothy Dozat, Natalia Silveira, Katri Haverinen, Filip Ginter, Joakim Nivre, and Christopher D. Manning. 2014. Universal Stanford dependencies: A cross-linguistic typology. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), pages 4585–4592, Reykjavik, Iceland. European Language Resources Association (ELRA). ↗
  3. Federica Gamba and Daniel Zeman. 2023. Universalising Latin Universal Dependencies: a harmonisation of Latin treebanks in UD. In Proceedings of the Sixth Workshop on Universal Dependencies (UDW, GURT/SyntaxFest 2023), pages 7–16, Washington, D.C. Association for Computational Linguistics. find
  4. Google DeepMind. 2026. Gemini 3 flash: A large multimodal model. https://deepmind.google/technologies/gemini/. Consulté le 10 mars 2026. ↗
  5. Federica Iurescia, Marco Passarotti, and Rachele Sprugnoli. 2026. Overview of the Dependency Parsing Task at EvaLatin 2026. In Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026), Palma, Mallorca (Spain). ELRA. find
  6. Milan Straka and Jana Straková. 2020. UD-Pipe at EvaLatin 2020: Contextualized embeddings and treebank embeddings. In Proceedings of LT4HALA 2020 - 1st Workshop on Language Technologies for Historical and Ancient Languages, pages 124–129, Marseille, France. European Language Resources Association (ELRA). find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/pommeret-2026-thivlvc-retrieval-augmented-dependency-parsing/