con.md

Constituency Parsing

This document provides examples training, evaluating and predicting with constituency parsers. All files passed through input arguments to load data are PTB bracket files (see docs/sample.ptb for an example).

Identifier	Parser	Paper	Arguments
con-idx	Absolute and relative indexing	Gómez-Rodríguez and Vilares (2018)	rel
con-tetra	Tetra-Tagging	Kitaev and Klein (2020)

Training

• Absolute indexing (Gómez-Rodríguez and Vilares, 2018 ) with XLNet (Yang et al., 2019) as encoder. Add --rel argument as <specific-args> to exchange absolute for relative positions.

python3 run.py con-idx -p results/con-idx-xlnet -c configs/xlnet.ini \
    train --train treebanks/ptb/train.ptb \
    --dev treebanks/ptb/dev.ptb \
    --test treebanks/ptb/test.ptb --num-workers 20

• Tetra-tagging (Kitaev and Klein, 2020) with XLNet (Yang et al., 2019) as encoder.

python3 run.py con-tetra -p results/con-tetra-xlnet -c configs/xlnet.ini \
    train --train treebanks/ptb/train.ptb \
    --dev treebanks/ptb/dev.ptb \
    --test treebanks/ptb/test.ptb --num-workers 20

Evaluation

Evaluate now the trained parser at results/con-tetra-xlnet/parser.pt with the same test file:

python3 run.py con-tetra -p results/con-tetra-xlnet/parser.pt eval treebanks/ptb/test.ptb --batch-size 50

Prediction

Predict with the trained parser at results/con-tetra-xlnet/parser.pt:

python3 run.py con-tetra -p results/con-tetra-xlnet/parser.pt predict \
    treebanks/ptb/test.ptb results/con-tetra-xlnet/pred.ptb --batch-size 50