Data Generation

This project includes utilities and scripts for automatic dataset generation. It is used in the following papers:

• Warstadt, A., Cao, Y., Grosu, I., Peng, W., Blix, H., Nie, Y., Alsop, A., Bordia, S., Liu, H., Parrish, A. and Wang, S.F., Bowman, S.R. 2019. Investigating BERT's Knowledge of Language: Five Analysis Methods with NPIs. arXiv preprint arXiv:1909.02597.
• Warstadt, A., Parrish, A., Liu, H., Mohananey, A., Peng, W., Wang, S. F., & Bowman, S. R. (2019). BLiMP: A Benchmark of Linguistic Minimal Pairs for English. arXiv preprint arXiv:1912.00582.
• Jeretic, P., Warstadt, A., Bhooshan, S., & Williams, A. (2020). Are Natural Language Inference Models IMPPRESsive? Learning IMPlicature and PRESupposition. arXiv preprint arXiv:2004.03066.

Usage

To run a sample data generation script, navigate to the data_generation directory and run the following command:

python -m generation_projects.blimp.adjunct_island

If all dependencies are present in your workspace, this will generate the adjunct_island dataset in BLiMP. Generation will take a minute to begin, after which point the progress can be watched in outputs/benchmark/adjunct_island.jsonl.

Branches

• With the exception of BLiMP, all project-specific code is kept in separate branches. BLiMP appears in master as a helpful examplar.
• Major branches include:
  • blimp
  • imppres
  • npi
  • msgs
  • structure_dependence qp

Project Structure

• The project contains the following packages:
  • generation_projects: scripts for generating data, organized into subdirectories by research project.
  • mturk_qc: code for carrying out Amazon mechanical turk quality control.
  • outputs: generated data, organized into subdirectories by research project.
  • results: experiment results files
  • results_processing: scripts for analyzing results and producing figures
  • utils: shared code for generation projects. Includes utilities for proecessing the vocabulary, generating constituents, manipulating generated strings, etc.
• It also contains a vocabulary file and documentation of the vocabulary:
  • vocabulary.csv: the vocab file.
  • vocab_documentation.md: the vocab documentation

Vocabulary

• The vocabulary file is vocabulary.csv.
• Each row in the .csv is a lexical item. Each column is feature encoding grammatical information about the lexical item. Detailed documentation of the columns can be found in vocab_documentation.md.
• The following notation is used to define selectional restrictions in the arg_1, arg_2, and arg_3 columns:

  <DISJUNCTION> := <CONJUNCTION> | <CONJUNCTION>;<DISJUNCTION>
  <CONJUNCTION> := <CONDITION> | <CONDITION>^<CONJUNCTION>
  <CONDITION> := <COLUMN>=<VALUE>
  

• In other words, the entire restriction is written in disjunctive normal form where ; is used for disjunction and ^ is used for conjunction.
• Example 1: arg_1 of lexical item breaking is animate=1. This means any noun appearing as the subject of breaking must have value 1 in the column animate.
• Example 2: arg_1 of lexical item buys is institution=1^sg=1;animate=1^sg=1. This means any noun appearing as the subject of breaking must meet one of the following conditions:
  1. have value 1 in column institution and value 1 in column sg, or
  2. have value 1 in column animate and value 1 in column sg.
• Disclaimer: As this project is under active development, data generated with different versions of the vocabulary may differ slightly.

Utils

• The utils package contains the shared code for the various generation projects.
  • utils.conjugate includes functions which conjugate verbs and add selecting auxiliaries/modals
  • utils.constituent_building includes functions which "do syntax". The following are especially useful:
    • verb_args_from_verb: gather all arguments of a verb into a dictionary
    • V_to_VP_mutate: given a verb, modify the expression to contain the string corresponding to a full VP
    • N_to_DP_mutate: given a noun, gather all arguments and a determiner, and modify the expression to contain the string corresponding to a full DP
  • utils.data_generator defines general classes that are instantianted by a particular generation project. The classes contain metadata fields, the main loop for a generating a dataset (generate_paradigm), and functions for logging and exception handling
  • utils.data_type contains the data_type necessary for the numpy structured array data structure used in the vocabulary.
    • if the columns of the vocabulary file are ever modified, this file must be modified to match.
  • utils.string_utils contains functions for cleaning up generated strings (removing extra whitespace, capitalization, etc.)
  • utils.vocab_sets contains constants for accessing commonly used sets of vocab entries. Building these constants takes about a minute at the beginning of running a generation script, but this speeds up generation of large datasets.
  • utils.vocab_table contains functions for creating and accessing the vocabulary table
    • get_all gathers all vocab items with a given restriction
    • get_all_conjunctive gathers all vocab items with the given restrictions

Citation

If you use the data generation project in your work, please cite the BLiMP paper:

@article{warstadt2019blimp,
  title={BLiMP: A Benchmark of Linguistic Minimal Pairs for English},
  author={Warstadt, Alex and Parrish, Alicia and Liu, Haokun and Mohananey, Anhad and Peng, Wei, and Wang, Sheng-Fu and Bowman, Samuel R},
  journal={arXiv preprint arXiv:1912.00582},
  year={2019}
}