On Explaining Your Explanations of BERT: An Empirical Study with Sequence Classification

Pytorch implmentation of Gradient Sensitivity, Graident X Input, Layerwise Revelance Propagation and Layerwise Attention Tracing for BERT-based models in just a single model.backward() call.

Citation

Zhengxuan Wu and Desmond C. Ong. 2020. On Explaining Your Explanations of BERT: An Empirical Study with Sequence Classification. Ms., Stanford University and National University of Singapore.

  @article{wu-ong-2020-explain,
    title={On Explaining Your Explanations of BERT: An Empirical Study with Sequence Classification},
    author={Wu, Zhengxuan and Ong, Desmond C.},
    journal={arXiv preprint},
    url={https://arxiv.org/abs/2101.00196},
    year={2020}}

Abstract

BERT, as one of the pretrianed language models, attracts the most attention in recent years for creating new benchmarks across NLP tasks via fine-tuning. One pressing issue is to open up the blackbox and explain the decision makings of BERT. A number of attribution techniques have been proposed to explain BERT models, but are often limited to sequence to sequence tasks. In this paper, we adapt existing attribution methods on explaining decision makings of BERT in sequence classification tasks. We conduct extensive analyses of four existing attribution methods by applying them to four different datasets in sentiment analysis. We compare the reliability and robustness of each method via various ablation studies. Furthermore, we test whether attribution methods explain generalized semantics across semantically similar tasks. Our work provides solid guidance for using attribution methods to explain decision makings of BERT for downstream classification tasks.

Introduction

Regular BERT training is for accuracy, but if you are interested in understanding why BERT made that decision, and which tokens contribute more towards that model decision, this repo will help you understand more about the BERT model. If you want to know the decision of your trained BERT, simple import our BERT model, and train with a gradient hook enabled. Your attribution scores will be calculated with a simple backward() call:

from BERT import *
# initialize model with our own BERT module
model = BERT()
# training loop
model.train()
# after training, simply call respective attribuiton method's backward function
model.backward()

Install Requirements

You will have to clone this repo, and install all the dependencies. You can skip this step if you have torch and cuda installed. That is all you need. You can also mannually install these without going through this installation headache that requirements.txt may give you.

cd BERT_LRP/code/
pip install -r requirements.txt

Download Pretrained BERT Model

You will have to download pretrained BERT model in order to execute the fine-tune pipeline. We recommand to use models provided by the official release on BERT from BERT-Base (Google's pre-trained models). Note that their model is in tensorflow format. To convert tensorflow model to pytorch model, you can use the helper script to do that. For example,

cd BERT_LRP/code/
python convert_tf_checkpoint_to_pytorch.py \
--tf_checkpoint_path uncased_L-12_H-768_A-12/bert_model.ckpt \
--bert_config_file uncased_L-12_H-768_A-12/bert_config.json \
--pytorch_dump_path uncased_L-12_H-768_A-12/pytorch_model.bin

BERT Model and Pretrain

Our BERT model is adapted from huggingface BERT model for text classification. If you want to take a look at the original model please search for BertForSequenceClassification. If you want to fine-tune or train a BERT classifier, you can either use their pipeline line or ours. Ours is tested against theirs before publishing. It is bug-free. To train a model, you can do something like this,

cd code/
CUDA_VISIBLE_DEVICES=0.1,2 python run_classifier.py \
--model_type BERTPretrain \
--eval_test \
--do_lower_case \
--max_seq_length 512 \
--train_batch_size 8 \
--eval_batch_size 8 \
--learning_rate 2e-5 \
--num_train_epochs 3 \
--seed 123 \
--task_name SST5 \
--data_dir ../datasets/SST5/ \
--vocab_file ../models/BERT-Google/vocab.txt \
--bert_config_file ../models/BERT-Google/bert_config.json \
--output_dir ../results/SST5-NewSeed/ \
--init_checkpoint ../models/BERT-Google/pytorch_model.bin

Take a look at code/util/processor.py to see how we process different datasets. We currently supports almost 10 different dataset loadings. You can create your own within 1 minute for loading data. You can specify your directories info above in the command.

Analyze, Attribution, Relevance and More

Once you have your model ready, save it to a location that you know (e.g., ../results/SST5/checkpoint.bin). Our example code how to get relevance scores is in a jupyter notebook format, which is much easier to read. This is how you will open it,

cd code/notebook/
jupyter notebook

Inside lrp_visualize, we provide an example on how to get relevance scores! In short, it is really easy, when you create your BERT model, just provide an extra argument which enables a variety of gradient hooks, by setting init_lrp=True. And then, in your evaluation loop, you can call model.backward_lrp(relevance_score).

fel-thomas/BERT_LRP