Parsing UD with <3 Matrices
This archive contains implementations of the methods from the papers:
- "Probing for Labeled Dependency Trees" (Müller-Eberstein, van der Goot and Plank, 2022a; ACL 2022);
- "Sort by Structure: Language Model Ranking as Dependency Probing" (Müller-Eberstein, van der Goot and Plank, 2022b; NAACL 2022).
After installing the required packages, and downloading the Universal Dependencies data, the experiments can be re-run using the
run-naacl.sh scripts respectively. Please see the instructions below for details.
The current codebase is compatible with both the ACL 2022 and NAACL 2022 experiments. However, if you wish to use the original version of the code used in the ACL paper, the appropriate commit is tagged
acl-2022 and can be found here.
This repository uses Python 3.6+ and the associated packages listed in the
(venv) $ pip install -r requirements.txt
The experiments are run on treebanks from Universal Dependencies version 2.8 (Zeman et al., 2021) which can be obtained separately at: https://lindat.mff.cuni.cz/repository/xmlui/handle/11234/1-3687.
The 13 target treebanks used are: AR-PADT (Hajič et al., 2009), EN-EWT (Silveira et al., 2014), EU-BDT (Aranzabe et al., 2015), FI-TDT (Pyysalo et al., 2015), HE-HTB (McDonald et al., 2013), HI-HDTB (Palmer et al., 2009), IT-ISDT (Bosco et al., 2014), JA-GSD (Asahara et al., 2018), KO-GSD (Chun et al., 2018), RU-SynTagRus (Droganova et al., 2018), SV-Talbanken (McDonald et al., 2013), TR-IMST (Sulubacak et al., 2016), ZH-GSD (Shen et al., 2016).
The nine target treebanks used are: AR-PADT (Hajič et al., 2009), EN-EWT (Silveira et al., 2014), FI-TDT (Pyysalo et al., 2015), GRC-PROIEL (Eckhoff et al., 2018), HE-HTB (McDonald et al., 2013a), KO-GSD (Chun et al., 2018), RU-GSD (McDonald et al., 2013b), SV-Talbanken (McDonald et al., 2013a), ZH-GSD (Shen et al., 2016).
Given that the pre-requesites are fulfilled, the experiments can be re-run using the respective
run-naacl.sh script. Please update the
exp_path variables within the scripts to point to the appropriate data and experiment directories on your machine.
(venv) $ ./run-acl.sh
will prepare the UD treebanks and train DepProbe as well as DirProbe models on all 13 target treebanks. Each experiment directory will contain predictions for the in-language test split and cross-lingual transfer results for the remaining 12 languages.
(venv) $ ./run-naacl.sh
will run all experiments using DepProbe from the main paper. This involves three random initializations across nine target languages with 4–6 language models each, totalling 138 setups.
We compare DepProbe and DirProbe with a full biaffine attention parser (BAP) as implemented in the MaChAmp framework version 0.3 (van der Goot et al., 2021). It is a freely available AllenNLP-based toolkit which is available at https://github.com/machamp-nlp/machamp.
For reproducing the results of BAP, we provide the parameter configurations and scripts for our dependency parsing experiments in
parsing/naacl/ respectively. Please use the same target treebanks in the probing experiments.
Parameters follow the default recommended configuration, but have been explicitly stated in order to ensure replicability. They can be found under
rs-42 for example denoting the use of the random seed 42. All other parameters are identical.
eval/langsim.py script can be used to calculate the cosine similarities between all language pairs using Lang2Vec (Littell et al., 2017).
(venv) $ python eval/langsim.py "ara" "eng" "eus" "fin" "heb" "hin" "ita" "jpn" "kor" "rus" "swe" "tur" "zho"
To perform an analysis of probe subspace angles across experiments/languages, use
(venv) $ python eval/ssa.py /path/to/exp1 /path/to/exp2 ...