la_core_web_trf: Latin NLP pipeline for spaCy. Trained on harmonized Universal Dependencies treebanks (Perseus, PROIEL, ITTB, LLCT, UDante, CIRCSE) with LASLA data. Includes POS tagging, morphological analysis, lemmatization, dependency parsing, named entity recognition, and sentence segmentation.

_{« Catus » — from "De cato seu fele" in Conrad Gesner's Historiae animalium liber primus de quadrupedibus viviparis (1551), p. 318}

Feature	Description
Name	`la_core_web_trf`
Version	`3.9.9`
spaCy	`>=3.8.11,<3.9.0`
Default Pipeline	`enclitic_splitter`, `transformer`, `senter`, `token_fix`, `normer`, `tagger`, `morphologizer`, `trainable_lemmatizer`, `lookup_lemmatizer`, `uv_normalizer`, `parser`, `harmonizer`, `remorpher`, `ner`, `trf_vectors`
Components	`enclitic_splitter`, `transformer`, `senter`, `token_fix`, `normer`, `tagger`, `morphologizer`, `trainable_lemmatizer`, `lookup_lemmatizer`, `uv_normalizer`, `parser`, `harmonizer`, `remorpher`, `ner`, `trf_vectors`
Vectors	0 keys, 0 unique vectors (0 dimensions)
Sources	CIRCSE/LASLA: LASLA Corpus LatinCy NER UD_Latin-CIRCSE UD_Latin-ITTB UD_Latin-LLCT UD_Latin-Perseus UD_Latin-PROIEL UD_Latin-UDante
License	`MIT`
Author	Patrick J. Burns
Contributors	Tim Geelhaar (annotation, error analysis [v3.5.2: morphologizer, tagger, parser]); Nora Bernhardt (NER); Vincent Koch (NER)

What's new in 3.9.9

Retrained la_core_web_trf only (sm/md/lg stay at 3.9.8).

Fixed Latin BERT tokenization. Backbone latincy/latin-bert v1.1.1 → v1.2.0. 3.9.1–3.9.8 inserted the subtokens \ 32 ;_ between every pair of words. Weights unchanged; only the tokenizer changed.
No remote code. Pure tokenizer.json; trust_remote_code = false.
UD test, 3.9.8 → 3.9.9: tag 94.13 → 94.42, POS 97.31 → 97.38, morph 93.95 → 94.28, lemma 97.20 → 97.39, UAS 86.88 → 87.56, LAS 82.01 → 82.69. NER dev F 91.06 → 91.95.

Known issues: see CHANGELOG.md.

What's new in 3.9.8

This is a maintenance release across the LatinCy pipeline family (la_core_web_sm/md/lg/trf). Model weights are unchanged from v3.9.x — no retraining. It also carries the 3.9.7 fix, which was built but never published. Evaluation metrics are not re-measured: the current test set (treebanks 3.10.13) differs from the one these weights were scored on, so the meta.json figures remain the v3.9 measurement.

Fixed empty lemmas. The edit-tree lemmatizer can select a delete-everything tree and emit "", which CSV/pandas consumers read back as NaN. The lookup lemmatizer now floors an empty lemma at the lowercased surface form. This model emitted "" for es and i on the dev corpus.
Fixed the sentence-boundary cascade after ): (3.9.7). token_fix's bracket balancing matched whole tokens, so a closer glued to trailing punctuation (): is a single token) was missed, and suppressing one boundary widened the scan for the next — collapsing the rest of a parenthetical passage into one sentence. Bracket depth is now counted per character and the scan reads a snapshot of the original boundaries.
Dependency pins. latincy-preprocess>=0.2.0,<0.6.0 (0.4.0–0.5.1 are Greek-only releases; Latin behavior unchanged) and la-latincy-lookups>=1.0.0,<1.1.0 — an upper bound is now declared, because 1.2.x tables regress lemmatization and must not resolve. The duplicate spacy requirement is removed.