This repository accompanies the paper (arXiv). Accepted to ACL 2020.

tl;dr read intro and one of the usage headings.

Tools to understand neural representations, and application to contextualizers. A “contexualizer” is a model producing a context-dependent word embedding.

Concretely, similarity measures, eg.

• CKA
• SVCCA

applied to SOTA contextualizers, eg.

• BERT
• ELMo

All similarity measures can be found in corr_methods.py.

We also experimented with novel attention-based similarity measures in attention_corr_methods.py.

This repository should be on your Python path.

export PYTHONPATH="${PWD}:${PYTHONPATH}"

The main script is main.py.

main.py [--methods [METHODS ...]] REPRESENTATION_FILES OUTPUT_FILE

For examples, see slurm (eg. mk_resultsN.sh and mk_resultsN-helper.sh). To see all options, run python main.py --help. Note that REPRESENTATION_FILES is a file containing an input file on each line. OUTPUT_FILE is a pickle dump.

main_attn.py is analogous.

You can also call the correlation methods directly from python. See ex.ipynb.

• var.py. Stuff you might want to change if you use this. eg, it has function fname2mname (filename to model name) that transforms /data/sls/temp/belinkov/contextual-corr-analysis/contextualizers/bert_large_cased/ptb_pos_dev.hdf5 to bert_large_cased-ptb_pos_dev.hdf5.
• analysis. Data analysis. The results that will be presented. analysis-n analyzes the result of experiment n.
• hnb. “Helper notebook.” Files in this directory are to
  • help me code
  • help the reader understand

  the resulting .py files.

  These are files containing a copy of the function with loops and co. destructured (run once with an arbitrary value, to help debugging).

  It may help you understand a function.

• slurm. SLURM scripts.

  Run directly as SCRIPTNAME.

• other. Everything else. Lots of junk.

Our pipeline is:

1. Generate representations (hdf5 files)
2. Run main.py on them
   1. Loads the representations (load_representations in corr_methods.py)
   2. Compute the correlations using the given methods
   3. Writes them to OUTPUT_FILE.
3. Analyze results (the OUTPUT_FILE above) in the analysis directory

New correlation methods should extend corr_methods.Method.

SVCCA similarities (Method CCA in corr_methods.py):

CKA attention similarity (Method AttnLinCKA in attention_corr_methods.py):