dep.md

Dependency Parsing

This document provides examples training, evaluating and predicting with dependency parsers. All files passed through input arguments to load data are CoNLL files.

Identifier	Parser	Paper	Arguments
dep-idx	Absolute and relative indexing	Strzyz et al. (2019)	rel
dep-pos	PoS-tag relative indexing	Strzyz et al. (2019)	gold
dep-bracket	Bracketing encoding ($k$-planar)	Strzyz et al. (2020)	k
dep-bit4	$4$-bit projective encoding	Gómez-Rodríguez et al. (2023)	proj
dep-bit7	$7$-bit $2$-planar encoding	Gómez-Rodríguez et al. (2023)
dep-eager	Arc-Eager system	Nivre and Fernández-González (2002)	stack, buffer, proj
dep-biaffine	Biaffine dependency parser	Dozat et al. (2016)
dep-hexa	Hexa-Tagging	Amini et al. (2023)	proj

Training

• Absolute indexing (Strzyz et al., 2019) with XLNet (Yang et al., 2019) as encoder. Add --rel argument as <specific-args> to exchange absolute for relative positions.

python3 run.py dep-idx -p results/dep-idx-xlnet -c configs/xlnet.ini \
    train --train treebanks/english-ewt/train.conllu \
    --dev treebanks/english-ewt/dev.conllu \
    --test treebanks/english-ewt/test.conllu --num-workers 20

• PoS-based relative indexing (Strzyz et al., 2019) with BiLSTMs as encoder. Add --gold argument as <specific-args> to not predict the PoS-tags but use the gold annotations.

python3 run.py dep-pos -p results/dep-pos-bilstm -c configs/bilstm.ini \
    train --train treebanks/english-ewt/train.conllu \
    --dev treebanks/english-ewt/dev.conllu \
    --test treebanks/english-ewt/test.conllu --num-workers 20

• Bracketing encoding (Strzyz et al., 2020) with $k=2$ with XLM (Conneau et al., 2019) as encoder:

python3 run.py dep-bracket -k 2 -p results/dep-bracket-xlm -c configs/xlm.ini \
    train --train treebanks/english-ewt/train.conllu \
    --dev treebanks/english-ewt/dev.conllu \
    --test treebanks/english-ewt/test.conllu --num-workers 20

• Arc-Eager transition-based system (Nivre and Fernández-González, 2002) where each state is represented with 1 position of the stack and 2 positions of the buffer. Use XLM (Conneau et al., 2019) as encoder:

python3 run.py dep-eager --stack 1 --buffer 2 -p results/dep-eager-xlm -c configs/xlm.ini \
    train --train treebanks/english-ewt/train.conllu \
    --dev treebanks/english-ewt/dev.conllu \
    --test treebanks/english-ewt/test.conllu --num-workers 20

• Hexa-Tagging (Amini et al., 2023) with XLNet as encoder and the head pseudo-projective transformation (modes available are head, head+path, path) from Nivre and Nilsson (2005):

python3 run.py dep-hexa -p results/dep-hexa-xlnet -c configs/xlnet.ini --proj head \
    --train treebanks/english-ewt/train.conllu \
    --dev treebanks/english-ewt/dev.conllu \
    --test treebanks/english-ewt/test.conllu --num-workers 20

Evaluation

Evaluate now the trained parser at results/dep-bracket-xlm/parser.pt with the same test file:

python3 run.py dep-bracket -p results/dep-bracket-xlm/parser.pt eval treebanks/english-ewt/test.conllu --batch-size 50

Prediction

Predict with the trained parser at results/dep-hexa-xlnet:

python3 run.py dep-hexa -p results/dep-hexa-xlnet/parser.pt predict \
     treebanks/english-ewt/test.conllu results/dep-hexa-xlnet/pred.conllu --batch-size 50