LM diagnostics

This repository contains the diagnostic datasets and experimental code for What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models, by Allyson Ettinger.

Diagnostic test data

The datasets folder contains TSV files with data for each diagnostic test, along with explanatory README files for each dataset.

Code

[All code now updated to be run with Python 3.]

The code in this section can be used to process the diagnostic datasets for input to a language model, and then to run the diagnostic tests on that language model's predictions. The code should be used in three steps:

Step 1: Process datasets to produce inputs for LM

proc_datasets.py can be used to process the provided datasets into 1) <testname>-contextlist files containing contexts (one per line) on which the LM's predictions should be conditioned, and b) <testname>-targetlist files containing target words (one per line, aligned with the contexts in *-contextlist) for which you will need probabilities conditioned on the corresponding contexts. Repeats in *-contextlist are intentional, to align with the targets in *-targetlist.

Basic usage:

python proc_datasets.py \
  --outputdir <location for output files> \
  --role_stim datasets/ROLE-88/ROLE-88.tsv \
  --negnat_stim datasets/NEG-88/NEG-88-NAT.tsv \
  --negsimp_stim datasets/NEG-88/NEG-88-SIMP.tsv \
  --cprag_stim datasets/CPRAG-34/CPRAG-34.tsv \
  --add_mask_tok

• add_mask_tok flag will append '[MASK]' to the contexts in *-contextlist, for use with BERT.
• <testname> comes from the following list: cprag, role, negsimp, negnat for CPRAG-34, ROLE-88, NEG-88-SIMP and NEG-88-NAT, respectively.

Step 2: Get LM predictions/probabilities

You will need to produce two files: one containing top word predictions conditioned on each context, and one containing the probabilities for each target word conditioned on its corresponding context.

Predictions: Model word predictions should be written to a file with naming modelpreds-<testname>-<modelname>. Each line of this file should contain the top word predictions conditioned on the context in the corresponding line in *-contextlist. Word predictions on a given line should be separated by whitespace. Number of predictions per line should be no less than the highest k that you want to use for accuracy tests.

Probabilities Model target probabilities should be written to a file with naming modeltgtprobs-<testname>-<modelname>. Each line of this file should contain the probability of the target word on the corresponding line of *-targetlist, conditioned on the context on the corresponding line of *-contextlist.

• <testname> list is as above. <modelname> should be the name of the model that will be input to the code in Step 3.

Step 3: Run accuracy and sensitivity tests for each diagnostic

prediction_accuracy_tests.py takes modelpreds-<testname>-<modelname> as input and runs word prediction accuracy tests.

Basic usage:

python prediction_accuracy_tests.py \
  --preddir <location of modelpreds-<testname>-<modelname>> \
  --resultsdir <location for results files> \
  --models <names of models to be tested, e.g., bert-base-uncased bert-large-uncased> \
  --k_values <list of k values to be tested, e.g., 1 5> \
  --role_stim datasets/ROLE-88/ROLE-88.tsv \
  --negnat_stim datasets/NEG-88/NEG-88-NAT.tsv \
  --negsimp_stim datasets/NEG-88/NEG-88-SIMP.tsv \
  --cprag_stim datasets/CPRAG-34/CPRAG-34.tsv

sensitivity_tests.py takes modeltgtprobs-<testname>-<modelname> as input and runs sensitivity tests.

Basic usage:

python sensitivity_tests.py \
  --probdir <location of modelpreds-<testname>-<modelname>> \
  --resultsdir <location for results files> \
  --models <names of models to be tested, e.g., bert-base-uncased bert-large-uncased> \
  --role_stim datasets/ROLE-88/ROLE-88.tsv \
  --negnat_stim datasets/NEG-88/NEG-88-NAT.tsv \
  --negsimp_stim datasets/NEG-88/NEG-88-SIMP.tsv \
  --cprag_stim datasets/CPRAG-34/CPRAG-34.tsv

Experimental code

run_diagnostics_bert.py is the code that was used for the experiments on BERT_BASE and BERT_LARGE reported in the paper, including perturbations.

Example usage:

python run_diagnostics_bert.py \
  --cprag_stim datasets/CPRAG-34/CPRAG-34.tsv \
  --role_stim datasets/ROLE-88/ROLE-88.tsv \
  --negnat_stim datasets/NEG-88/NEG-88-NAT.tsv \
  --negsimp_stim datasets/NEG-88/NEG-88-SIMP.tsv \
  --resultsdir <location for results files> \
  --bertbase <BERT BASE location> \
  --bertlarge <BERT LARGE location> \
  --incl_perturb

• bertbase and bertlarge specify locations for PyTorch BERT_BASE and BERT_LARGE models -- each folder is expected to include vocab.txt, bert_config.json, and pytorch_model.bin for the corresponding PyTorch BERT model. (Note that experiments were run with the original pytorch-pretrained-bert version, so I can't guarantee identical results with the updated pytorch-transformers.)
• incl_perturb runs experiments with all perturbations reported in the paper. Without this flag, only runs experiments without perturbations.