BERT-Syntax

Assesing the syntactic abilities of BERT.

What

Evaluate Google's BERT-Base and BERT-Large models on the syntactic agreement datasets from Linzen, Goldberg and Dupoux 2016 and Marvin and Linzen 2018 and Gulordava et al 2018.

This is quite messy, as I hacked it together between things here and there. But I also believe it is accurate. This lists the data files and shows how to run the evaluation. For more details and results, see the arxiv report.

Data Files

Data taken from the github repos of Linzen, Goldberg and Dupoux (LGD), Marvin and Linzen (ML), and Gulordava et al.

File	Description
marvin_linzen_dataset.tsv	stimuli from Marvin and Linzen. I dumped it from the pickle files of ML
wiki.vocab	from LGD, used for verb inflections (wiki.vocab)
lgd_dataset.tsv	processed data from LGD
generated.tab	data from Gulordava et al (generated.tab)

lgd_dataset.tsv is created by

wget http://tallinzen.net/media/rnn_agreement/agr_50_mostcommon_10K.tsv.gz
gunzip agr_50_mostcommon_10K.tsv.gz
python make_linzen_goldberg_testset.py > lgd_dataset.tsv

Obtaining the results

pip install pytorch_pretrained_bert

python eval_bert.py > results/lgd_results_large.txt
python eval_bert.py base > results/lgd_results_base.txt
python eval_bert.py marvin > results/marvin_results_large.txt
python eval_bert.py marvin base > results/marvin_results_base.txt
python eval_bert.py gul > results/gulordava_results_large.txt
python eval_bert.py gul base > results/gulordava_results_base.txt

Generating tables (for the PDF)

python gen_marvin_tbl.py 
python gen_lgd_tbl.py
python gen_gul_tbl.py