Repository for LASLA Latin models: the models were fine-tuned by Thibault Clérice, data are based on LASLA data but some adaptation might be found.

Check latest release, under assets

Note: the model is currently being fine-tuned in the context of my PhD. I'll fill this part when it will be done.

The training set is roughly 1.5M tokens, dev test roughly 10k and test 169822. This is not counting punctuation, as LASLA data are lacking punctuation.

• Enclitics are kept in a single token
  • Enclitic lemma are separated as such token[Caesarque] == lemma[Caesar界que]
  • Morphology is the morphology of the first token
• Only numbers 1, 2 and 3 are known. Roman numbers are unknown.
• All punctuation signs are unknown, including the one used in abbr. token[C] == lemma[Gaius]
• Lemma and tokens now accept lower and uppercasing. Noise was introduced in the dataset for better results.

More details:

• lemma
• pos
• Gend
• Numb
• Case
• Deg
• Mood_Tense_Voice
• Person
• Dis

Definition to come. POS has no NOMcom vs. NOMpro (needs to be fixed in a later training.)

The model LASLA+ is trained on additionnal data, creating some noise in the original dataset and resulting in apparently worse results on classical data (approxim. -0.3%). It's results are detailed in LASLA-plus.md.

More details:

• lemma
• pos
• Gend
• Numb
• Case
• Deg
• Mood_Tense_Voice
• Person
• Dis

Definition to come

• D. Longrée, C. Philippart de Foy & G. Purnelle. « Structures phrastiques et analyse automatique des données morphosyntaxiques : le projet LatSynt », in S. Bolasco, I. Chiari & L. Giuliano (eds), Statistical Analysis of Textual Data, Proceedings of 10th International Conference Journées d'Analyse statistique des Données Textuelles, 9-11 June 2010, Sapienza University of Rome, Rome, LED, pp. 433-442.
• D. Longrée & C. Poudat, « New Ways of Lemmatizing and Tagging Classical and post-Classical Latin: the LATLEM project of the LASLA », in P. Anreiter & M. Kienpointner (éd.), Proceedings of the 15th International Colloquium on Latin Linguistics, (Innsbrucker Beiträge zur Sprachwissenschaft), Innsbruck, 2010, pp. 683-694.
• D. Longrée & C. Philippart de Foy & G. Purnelle, « Subordinate clause boundaries and word order in Latin: the contribution of the L.A.S.L.A. syntactic parser project LatSynt », in P. Anreiter & M. Kienpointner, éd.), Proceedings of the 15th International Colloquium on Latin Linguistics, (Innsbrucker Beiträge zur Sprachwissenschaft), Innsbruck, 2010, pp. 673-681.
• D. Longrée & Poudat C., « Variations langagières et annotation morphosyntaxique du latin classique », TAL, 50 – n° 2/2009, Special issue on "Natural Language Processing and Ancient Languages", pp. 129-148.
• E. Manjavacas & Á. Kádár & M. Kestemont, « Improving Lemmatization of Non-Standard Languages with Joint Learning », Proceedings of the 2019 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Special issue on "Natural Language Processing and Ancient Languages", 2019, pp. 493--1503.
• Enrique Manjavacas & Mike Kestemont. (2019, January 17). emanjavacas/pie v0.1.3 (Version v0.1.3). Zenodo. http://doi.org/10.5281/zenodo.2542537 Check the latest version here :Zenodo DOI
• Thibault Clérice. (2020, April 28). PonteIneptique/latin-lasla-models: 0.0.0 (Version 0.0.0). Zenodo. http://doi.org/10.5281/zenodo.3773328

The web application and its maintenance is done by Thibault Clérice ( @ponteineptique ). To learn how to cite this repository, go check our releases.

The model is based on the LASLA data.