Does BERT Rediscover a Classical NLP Pipeline?

Code, data, and results of the COLING 2022 paper Does BERT Rediscover a Classical NLP Pipeline?

Abstract: Does BERT store surface knowledge in its bottom layers, syntactic knowledge in its middle layers, and semantic knowledge in its upper layers? In re-examining Jawahar et al. (2019) and Tenney et al.'s (2019) probes into the structure of BERT, we have found that the pipeline-like separation that they were seeking lacks conclusive empirical support. BERT's structure is, however, linguistically grounded, although perhaps in a way that is more nuanced than can be explained by layers alone. We introduce a novel probe, called GridLoc, through which we can also take into account token positions, training rounds, and random seeds. Using GridLoc, we are able to detect other, stronger regularities that suggest that pseudo-cognitive appeals to layer depth may not be the preferred mode of explanation for BERT's inner workings.

Plots and Results

The plots and results used in the paper are available at https://doi.org/10.5683/SP3/PCZHN4. This Dataverse repository should contain the following files:

gridloc_checkpoints.tar.gz[1.8GB] contains all the probe model checkpoints and training logs.
gridloc_plots.tar.gz[2.5GB] contains all the plots generated for the paper.

The heat map visualisation of the layer performance probing result of Jawahar et al. (2019) (figure 2 of the paper), and the architecture of GirdLoc (figure 3 of the paper) are available at plots/others.

Quick Start

Install Dependencies

Create a virtual environment and install the required dependencies.

virtualenv env -p python3
source env/bin/activate
pip install -r requirements.txt

Install PyTorch. Please follow PyTorch's official installation guide. This paper is implemented using PyTorch version 1.10.2+cu102.

Install Stanza English model. (Required only for TreeDepth analysis.)

python -c "import stanza; stanza.download('en')"

Data Preparation

Get SentEval data.

mkdir data
cd data/
git clone https://github.com/facebookresearch/SentEval
cd ..

Train Probes

Initiate the training process by running the script scripts/train_probes.py.

python scripts/train_probes.py

Plot Layer Attention Weight

The script scripts/plot_layer_weight_centers.py contains the code to compute the layer attention weight distributions used in section 5.1 and section 5.2 of the paper.

python scripts/plot_layer_weight_centers.py

The script will generate 3 files for each epoch of the probe: an .svg file, a .csv file, and an .npy file.

layer_weight_center_epoch_X.svg: The vector file of the layer attention weights plot.
layer_weight_center_epoch_X.csv: The exact attention weights of the 12 layers.
layer_weight_center_epoch_X.npy: The serialised NumPy file containing the layer attention weights of every sentence in the task's test set. See scripts/read_npy.py for more details on how to load and read the results.

Similarly, the script scripts/plot_top_layer_distribution.py contains the code to compute the top layer distributions used in section 5.1 and section 5.2 of the paper.

python scripts/plot_layer_weight_centers.py

The script will only generate a top_layer_distribution_X.svg file containing the plot of the top layer distributions.

Plot Sentences

The script scripts/plot_layer_weight_centers.py contains the code to plot token-position attention weights of sentences. These plots are used in section 5.3 and section 5.4 of the paper.

python scripts/plot_sentences.py

The script will generate a sentence_plot_XXXXXX.svg file containing the token-position attention weights heat map. The sentence ID in the file name directly corresponds to the line number of SentEval data.

TreeDepth Correlation Analysis

The script scripts/tree_depth_analysis.py contains the code to compute the average layer of different POSs and the correlation analysis of section 5.4 of the paper.

python scripts/tree_depth_analysis.py

NOTE: You may need to change the paths in the scripts to match your directory organisation.

Citation

@inproceedings{niu-etal-2022-bert,
    title = "Does {BERT} Rediscover a Classical {NLP} Pipeline?",
    author = "Niu, Jingcheng  and
      Lu, Wenjie  and
      Penn, Gerald",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.278",
    pages = "3143--3153"
}

Contact

Email: {niu,luwenjie,gpenn}@cs.toronto.edu

frankniujc/gridloc_probe