mcoavoux/self-attentive-parser

High-accuracy NLP parser with models for 11 languages.

This is a fork from Nikita Kitaev's self-attentive parser that was used to perform constituency parsing experiments described in FlauBERT paper.

Installation: Python / Conda set-up

git clone https://github.com/mcoavoux/self-attentive-parser
cd self-attentive-parser

conda create --name FlaubertParse --file requirements-conda.txt
conda activate FlaubertParse
pip install -r requirements.txt

Parse French sentences with pretrained models

Download and uncompress pretrained models:

wget https://zenodo.org/record/3696502/files/french_parsing_models.tar.gz
mkdir french_parsing_models
tar xzf french_parsing_models.tar.gz -C french_parsing_models

The folder french_parsing_models will contain 3 pretrained models (trained respectively with flaubert-base, flaubert-large, camembert-base).

Parse with pretrained model:

conda activate FlaubertParse

# Parse with Flaubert Pretrained model (adapt batch size depending on available memory)
python src/main.py parse --model-path-base french_parsing_models/model_flaubert_base_seed1_dev=88.96.pt --input-path to_parse.txt --output-path to_parse.parsed --eval-batch-size 10

# Parse with Camembert Pretrained model
python src/main.py parse --model-path-base french_parsing_models/model_camembert_seed2_dev=88.35.pt --input-path to_parse.txt --output-path to_parse.parsed --eval-batch-size 10

The input file should contain one sentence per line (already tokenized as in the French Treebank). You can check if the results are correct by comparing the models' output to to_parse_{camembert,flaubert}_expected_result.parsed.

Retrain and evaluate models

If you want to retrain with Fasttext embeddings, you should download them here.

1. Install Evalb

    cd self-attentive-parser
    cd EVALB_SPMRL
    make
    cd ..
   

2. Train models

    cd flaubert
   
    # change absolute path to conda environment in flaubert/expe_master.sh (first line)
    # change absolute path to data folder in flaubert/expe_master.sh (`SPMRL` variable)
    bash oar_expe.sh
    
    # in case oar scheduler is not available, just run commands inside the string passed in arguments of oarsub in oar_expe.sh
   

Parse with models

SPMRL=<path to SPMRL_SHARED_2014_NO_ARABIC>
testpath="${SPMRL}/FRENCH_SPMRL/gold/ptb/test/test.French.gold.ptb"
devpath="${SPMRL}/FRENCH_SPMRL/gold/ptb/dev/dev.French.gold.ptb"

model=<path to trained model>

python src/main.py test --test-path ${testpath} --evalb-dir EVALB_SPMRL --model-path-base ${model} > log_eval_test
python src/main.py test --test-path ${devpath}  --evalb-dir EVALB_SPMRL --model-path-base ${model} > log_eval_dev


# To parse from ensemble of several models (i.e. fine-tuned CamemBERT + fine-tuned FlauBERT)
models="<path to fine-tuned camembert> <path to fine-tuned flaubert>"
python src/main.py ensemble --test-path ${testpath} --evalb-dir EVALB_SPMRL --model-path-base ${models} --eval-batch-size 50 > log_eval_test
python src/main.py ensemble --test-path ${devpath}  --evalb-dir EVALB_SPMRL --model-path-base ${models} --eval-batch-size 50 > log_eval_dev