cpllab / syntactic-generalization

Code and data for "A Systematic Assessment of Syntactic Generalization in Neural Language Models"

A Systematic Assessment of Syntactic Generalization in Neural Language Models

This repository contains the analysis code and raw data for "A Systematic Assessment of Syntactic Generalization in Neural Language Models", accepted to appear at ACL 2020. If you use any of our materials, we ask that you cite the paper as follows:

@inproceedings{Hu:et-al:2020,
  author = {Hu, Jennifer and Gauthier, Jon and Qian, Peng and Wilcox, Ethan and Levy, Roger},
  title = {A systematic assessment of syntactic generalization in neural language models},
  booktitle = {Proceedings of the Association of Computational Linguistics},
  year = {2020}
}

Figures

To reproduce the figures in our paper, please run the iPython notebook at notebooks/main.ipynb. The only dependencies are basic scientific Python (numpy, pandas, scipy, matplotlib, etc.).

The raw data can be found at data/raw.

Model parameters

Parameters for our trained models can be provided upon request. Please raise an issue or get in touch with Jenn ([e-mail]).

BLLIP datasets

Unfortunately, we cannot distribute the BLLIP data we used to train our models because it is copyright-protected. However, you may be able to obtain BLLIP from the LDC through your institution: https://catalog.ldc.upenn.edu/LDC2000T43

If you do have access to BLLIP, please read the following instructions for constructing our corpora and train-dev-test splits.

Dev and test

The dev and test splits are shared across our BLLIP corpora. They can be obtained as follows:

• Dev (shared across corpora): first 500 sentences of first section for each year (1500 total)
• Test (shared across corpora): first 1000 sentences of second section for each year (3000 total)

Training

Our training sets satisfy the following properties:

1. Each year from the full corpus (1987-1989) is equally represented.
2. Within each year, each section is equally represented (and exact section numbers are uniformly sampled).
3. Each corpus is a proper subset of the larger corpora.

The exact section numbers are as follows:

• BLLIP-LG: all sections except [1987_001-002, 1988_001-002, 1989_010-011]

• BLLIP-MD:

  • 30 sections of 1987: [5, 10, 18, 21, 22, 26, 32, 35, 43, 47, 48, 49, 51, 54, 55, 56, 57, 61, 62, 65, 71, 77, 79, 81, 90, 96, 100, 105, 122, 125]
  • 30 sections of 1988: [12, 13, 14, 17, 23, 24, 33, 39, 40, 47, 48, 54, 55, 59, 69, 72, 73, 76, 78, 79, 83, 84, 88, 89, 90, 93, 94, 96, 102, 107]
  • 30 sections of 1989: 012-041

• BLLIP-SM:

  • 10 sections of 1987: [35, 43, 48, 54, 61, 71, 77, 81, 96, 122]
  • 10 sections of 1988: [24, 54, 55, 59, 69, 73, 76, 79, 90, 107]
  • 10 sections of 1989: [12, 13, 15, 18, 21, 22, 28, 37, 38, 39]

• BLLIP-XS:

  • 2 sections of 1987: [71, 122]
  • 2 sections of 1988: [54, 107]
  • 2 sections of 1989: [28, 37]