This repository contains the code and data for the paper:

Measuring and Improving Consistency in Pretrained Language Models

as well as the resource: ParaRel 🤘

Since this work required running a lot of experiments, it is structured by scripts that automatically runs many sub-experiments, on parallel servers, and tracking using an experiment tracking website: wandb, which are then aggregated using a jupyter notebook. To run all the experiments I used task spooler, a queue-based software that allows to run multiple commands in parallel (and store the rest in a queue)

It is also possible to run individual experiments, for which one can look for in the corresponding script.

For any question, query regarding the code, or paper, please reach out at [email protected]

If you're only interested in the data, you can find it under data. Each file contains the paraphrases patterns for a specific relation, in a json file.

conda create -n pararel python=3.7 anaconda
conda activate pararel

pip install -r requirements.txt

add project to path:

export PYTHONPATH=${PYTHONPATH}:/path-to-project

In case you just want to start with the filtered data we used (filtering objects that consist more than a single word piece in the LMs we considered), you can find them here. Otherwise:

First, begin by downloading the trex dataset from here, alternatively, check out the LAMA github repo. Download it to the following folder so that the following folder would exist: data/trex/data/TREx along with the relevant files

Next, in case you want to rerun automatically some/all of the experiments, you will need to update the paths in the runs scripts with your folder path and virtual environment.

Filter data from trex, to include only triplets that appear in the inspected LMs in this work: bert-base-cased, roberta-base, albert-base-v2 (as well as the larger versions, that contain the same vocabulary)

python runs/pararel/filter.py

A single run looks like the following:

python lm_meaning/lm_entail/filter_data.py \
       --in_data data/trex/data/TREx/P106.jsonl \
       --model_names bert-base-cased,bert-large-cased,bert-large-cased-whole-word-masking,roberta-base,roberta-large,albert-base-v2,albert-xxlarge-v2 \
       --out_file data/trex_lms_vocab/P106.jsonl

Evaluate consistency:

python runs/eval/run_lm_consistent.py

A single run looks like the following:

python pararel/consistency/encode_consistency_probe.py \
       --data_file data/trex_lms_vocab/P106.jsonl \
       --lm bert-base-cased \
       --graph data/pattern_data/graphs/P106.graph \
       --gpu 0 \
       --wandb \
       --use_targets

Encode the patterns along with the subjects, to save the representations:

python runs/pararel/encode_text.py

A single run looks like the following:

python lm_meaning/encode/encode_text.py \
       --patterns_file data/pattern_data/graphs_json/P106.jsonl \
       --data_file data/trex_lms_vocab/P106.jsonl \
       --lm bert-base-cased \
       --pred_file data/output/representations/P106_bert-base-cased.npy \
       --wandb

The code and README are available here

Q: Why do you report 31 N-1 relations, whereas in the LAMA paper there are only 25?

A: Explanation

If you find this work relevant to yours, please cite us:

@article{Elazar2021MeasuringAI,
  title={Measuring and Improving Consistency in Pretrained Language Models},
  author={Yanai Elazar and Nora Kassner and Shauli Ravfogel and Abhilasha Ravichander and Ed Hovy and Hinrich Schutze and Yoav Goldberg},
  journal={ArXiv},
  year={2021},
  volume={abs/2102.01017}
}