TO REPRODUCE THE EXPERIMENT DESCRIBED IN THE PAPER (need a GPU):

1. Install the dependencies as per requirements.txt
2. python3 train_on_dataset.py path/to/traininig_dataset path/to/dev_dataset path/to/save/model
3. python3 evaluate_hf.py path/to/save/model/saved/ path/to/test_dataset path/to/log/output

FOR THE NCRF++ model:

1. Install the dependencies as per requirements.txt (in particular, torch==1.11.0)
2. Download the glove.840B.300d word embeddings from https://nlp.stanford.edu/projects/glove/
3. The data is provided under "data". It was extracted from the ERG treebanks 2020 release (see link in the paper).
4. Now simply run the (slightly modified) NCRF++ on the data: python3 main.py --config erg.tran-best.config
5. To decode, run: python3 main.py --config erg.decode.config

You can change the test data in the config to obtain numbers for each test corpus separately.

FOR THE BASELINE MODEL:

1. Prepare the data: 1.1. Download the ERG treebanks from svn checkout http://svn.delph-in.net/erg/tags/2020 1.2. Locate the treebanks under tsdb/gold/ 1.3. Prepare the data split as recommended in the file redwoods.xls (found in the same ERG archive). Put all the training corpora in a folder called "train", all the dev ones in "dev", and all the test ones in "test". You should have 57 corpora in train, 2 in dev, and 14 in test.

2. Put the ERG lexicons in a separate folder, you should have 4 files there: gle.tdl, lexicon-rbst.tdl, lexicon.tdl, and ple.tdl

3. Run letype_extractor as follows:

   python3 letype_extractor.py path-to-lexicons/ path-to-treebanks/ custom-name-for-run- nonautoreg

(Just use any name for your run, such as "paper-repro" or whatever. The "nonautoreg" flag is needed in some cases to train autoregressive vs nonautoregressive models; just put "nonautoreg" there. This step will produce a folder under "output/your-run-name" with data in it.

4. Now run the vectorizer: python3 vectorizer.py path-to-your-run-folder/ nonautoreg

This will create the vectorized data under the same folder.

5. Now train the models (this will train two models: SVM and MaxEnt OVR L1 SAGA:

python3 classic_classifiers.py train output/your-run-name/ nonautoreg

(the OVR MaxEnt model may take up to 48 hours to train, depending on the machine. SVM trains under a few hours).

5. To decode, for dev or test:

   python3 classic_classifiers.py test output/your-run-name/ dev nonautoreg

   or:

   python3 classic_classifiers.py test output/your-run-name/ test nonautoreg

Acknowledgments

We acknowledge the European Union's Horizon Europe Framework Programme which funded this research under the Marie Skłodowska-Curie postdoctoral fellowship grant HORIZON-MSCA‐2021‐PF‐01 (GAUSS, grant agreement No 101063104); the European Research Council (ERC), which has funded this research under the Horizon Europe research and innovation programme (SALSA, grant agreement No 101100615); Grant SCANNER-UDC (PID2020-113230RB-C21) funded by MICIU/AEI/10.13039/501100011033; Xunta de Galicia (ED431C 2020/11); and Centro de Investigación de Galicia ‘‘CITIC’’, funded by the Xunta de Galicia through the collaboration agreement between the Consellería de Cultura, Educación, Formación Profesional e Universidades and the Galician universities for the reinforcement of the research centres of the Galician University System (CIGUS). We also acknowledge grant GAP (PID2022-139308OA-I00) funded by MCIN/AEI/10.13039/501100011033/ and by ERDF ``A way of making Europe''.