latin-standardized-treebanks

Code and data for: Marisa Hudspeth, Brendan O’Connor, Laure Thompson, "Latin Treebanks in Review: An Evaluation of Morphological Tagging Across Time." Machine Learning for Ancient Languages (ML4AL) Workshop, 2024.

Directory Structure

├── code
│   ├── conversion_scripts

• Code for converting from UD/LASLA tagset to a Standard Latin tagset.

│   └── morph_tagging
│       └── scripts

• Code for finetuning LatinBERT (Bamman and Burns, 2020) for morphological tagging.

├── data
│   ├── converted_treebanks
│   │   ├── lasla
│   │   ├── ud
│   │   ├── full_ud_sets
│   │   ├── train_test_splits
│   │   └── split_ids.json

• Directories containing UD and LASLA treebanks converted to a Standard Latin tagset:
  • lasla: converted files from the LASLA repo
  • ud: converted files from 5 UD treebanks. Original, UD train-test splits
  • full_ud_sets: individual files for each unique text in the 5 UD treebanks. No train/test splits
  • train_test_splits: custom train/test splits proposed in our paper. Each unique work in LASLA and UD, has separate files for its train, dev, and test sets (whichever are applicable).
  • split_ids.json: for each unique work in LASLA and UD, lists which sentence IDs belong to which train/dev/test set

│   ├── original_treebanks 

• Directory to hold original treebanks (not included in repo, but can be downloaded)

└── └── metadata.csv

• For each unique work in LASLA and UD treebanks, lists: source treebank(s), time period, century, number of sentences. For UD texts, also includes genre labels.

Setup for converting treebanks

If you'd like to replicate the conversion process, follow these steps. We've also provided the converted files already in data/converted_treebanks/, so these steps are optional.

First, clone the Latin treebanks. The first is 5 harmonized UD treebanks by Gamba and Zeman (2023), and the second is the LASLA treebank (Denooz 2004).

cd data/original_treebanks/
git clone https://github.com/fjambe/Latin-variability.git
git clone https://github.com/CIRCSE/LASLA.git

Then run the conversion scripts inside latin-standardized-treebanks/code/conversion_scripts/:

• convert_ud_treebanks.py: creates two subdirectories, ud and lasla inside of data/converted_treebanks/, and converts treebanks to our tagset.
• make_custom_treebanks.py: splits the converted ud treebanks inside data/ud/ into files for each individual work, which are saved in data/full_ud_sets/. Also creates custom train/test splits, saved in data/train_test_splits.

Finetuning LatinBERT for morphological tagging

Clone the LatinBERT repo:

cd code/
git clone https://github.com/dbamman/latin-bert.git
mv morph_tagging ./latin-bert/case_studies/

Optional: generate the tagset (we have already included the tagset file in data/converted_treebanks/morph.tagset):

cd latin-bert/case_studies/morph_tagging/scripts
python generate_tagset.py -f ../../../../../data/converted_treebanks/train_test_splits/*.conllu > morph.tagset

Example train command:

python latin_sequence_labeling.py -m train \
    --bertPath ../../../models/latin_bert \
    --tokenizerPath ../../../models/subword_tokenizer_latin/latin.subword.encoder \
    -r [LIST OF TRAINING FILEPATHS] \
    -g ../../../../../data/converted_treebanks/morph.tagset \
    -f [MODEL SAVE PATH] \
    -z fscore \
    --max_epochs [MAX_EPOCHS] \
    --load_from_checkpoint_num=-1 \
    --save_every_n=-1 

Options inside brackets [] are for you to fill in.

Example predict command:

python latin_sequence_labeling.py -m predict \
    --bertPath ../../../models/latin_bert \
    --tokenizerPath ../../../models/subword_tokenizer_latin/latin.subword.encoder \
    -g ../../../../../data/converted_treebanks/morph.tagset \
    -f [MODEL SAVE PATH]\
    -z fscore \
    -i [LIST OF TEST FILEPATHS] \
    -o [LIST OF OUTPUT FILEPATHS, PARALLEL TO TEST PATHS]

Our code is adapted from the POS tagging case study code for the LatinBERT paper.