All datasets
BLiMP

BLiMP

The Benchmark of Linguistic Minimal Pairs, a challenge set for evaluating the linguistic knowledge of language models (LMs) on major grammatical phenomena in English, finds that state-of-the-art models identify morphological contrasts related to agreement reliably, but they struggle with some subtle semantic and syntactic phenomena.

cc-by-4.0
10k<n<100k
Text Classification
English
by @AIOZNetwork
•
3

Last updated: 20 days ago

Details
Files
Discussions
0

BLiMP

Summary

Introduction

BLiMP is a challenge set for evaluating what language models (LMs) know about major grammatical phenomena in English. BLiMP consists of 67 sub-datasets, each containing 1000 minimal pairs isolating specific contrasts in syntax, morphology, or semantics. The data is automatically generated according to expert-crafted grammars.

Dataset Structure

Data Instances

adjunct_island

  • Size of downloaded dataset files: 0.36 MB
  • Size of the generated dataset: 0.17 MB
  • Total amount of disk used: 0.52 MB

An example of 'train' looks as follows.

{
    "UID": "tough_vs_raising_1",
    "field": "syntax_semantics",
    "lexically_identical": false,
    "linguistics_term": "control_raising",
    "one_prefix_method": false,
    "pair_id": 2,
    "sentence_bad": "Benjamin's tutor was certain to boast about.",
    "sentence_good": "Benjamin's tutor was easy to boast about.",
    "simple_LM_method": true,
    "two_prefix_method": false
}

anaphor_gender_agreement

  • Size of downloaded dataset files: 0.44 MB
  • Size of the generated dataset: 0.14 MB
  • Total amount of disk used: 0.57 MB

An example of 'train' looks as follows.

{
    "UID": "tough_vs_raising_1",
    "field": "syntax_semantics",
    "lexically_identical": false,
    "linguistics_term": "control_raising",
    "one_prefix_method": false,
    "pair_id": 2,
    "sentence_bad": "Benjamin's tutor was certain to boast about.",
    "sentence_good": "Benjamin's tutor was easy to boast about.",
    "simple_LM_method": true,
    "two_prefix_method": false
}

anaphor_number_agreement

  • Size of downloaded dataset files: 0.45 MB
  • Size of the generated dataset: 0.14 MB
  • Total amount of disk used: 0.59 MB

An example of 'train' looks as follows.

{
    "UID": "tough_vs_raising_1",
    "field": "syntax_semantics",
    "lexically_identical": false,
    "linguistics_term": "control_raising",
    "one_prefix_method": false,
    "pair_id": 2,
    "sentence_bad": "Benjamin's tutor was certain to boast about.",
    "sentence_good": "Benjamin's tutor was easy to boast about.",
    "simple_LM_method": true,
    "two_prefix_method": false
}

animate_subject_passive

  • Size of downloaded dataset files: 0.46 MB
  • Size of the generated dataset: 0.15 MB
  • Total amount of disk used: 0.61 MB

An example of 'train' looks as follows.

{
    "UID": "tough_vs_raising_1",
    "field": "syntax_semantics",
    "lexically_identical": false,
    "linguistics_term": "control_raising",
    "one_prefix_method": false,
    "pair_id": 2,
    "sentence_bad": "Benjamin's tutor was certain to boast about.",
    "sentence_good": "Benjamin's tutor was easy to boast about.",
    "simple_LM_method": true,
    "two_prefix_method": false
}

animate_subject_trans

  • Size of downloaded dataset files: 0.43 MB
  • Size of the generated dataset: 0.13 MB
  • Total amount of disk used: 0.57 MB

An example of 'train' looks as follows.

{
    "UID": "tough_vs_raising_1",
    "field": "syntax_semantics",
    "lexically_identical": false,
    "linguistics_term": "control_raising",
    "one_prefix_method": false,
    "pair_id": 2,
    "sentence_bad": "Benjamin's tutor was certain to boast about.",
    "sentence_good": "Benjamin's tutor was easy to boast about.",
    "simple_LM_method": true,
    "two_prefix_method": false
}

Data Fields

The data fields are the same among all splits.

adjunct_island

  • sentence_good: a string feature.
  • sentence_bad: a string feature.
  • field: a string feature.
  • linguistics_term: a string feature.
  • UID: a string feature.
  • simple_LM_method: a bool feature.
  • one_prefix_method: a bool feature.
  • two_prefix_method: a bool feature.
  • lexically_identical: a bool feature.
  • pair_id: a int32 feature.

anaphor_gender_agreement

  • sentence_good: a string feature.
  • sentence_bad: a string feature.
  • field: a string feature.
  • linguistics_term: a string feature.
  • UID: a string feature.
  • simple_LM_method: a bool feature.
  • one_prefix_method: a bool feature.
  • two_prefix_method: a bool feature.
  • lexically_identical: a bool feature.
  • pair_id: a int32 feature.

anaphor_number_agreement

  • sentence_good: a string feature.
  • sentence_bad: a string feature.
  • field: a string feature.
  • linguistics_term: a string feature.
  • UID: a string feature.
  • simple_LM_method: a bool feature.
  • one_prefix_method: a bool feature.
  • two_prefix_method: a bool feature.
  • lexically_identical: a bool feature.
  • pair_id: a int32 feature.

animate_subject_passive

  • sentence_good: a string feature.
  • sentence_bad: a string feature.
  • field: a string feature.
  • linguistics_term: a string feature.
  • UID: a string feature.
  • simple_LM_method: a bool feature.
  • one_prefix_method: a bool feature.
  • two_prefix_method: a bool feature.
  • lexically_identical: a bool feature.
  • pair_id: a int32 feature.

animate_subject_trans

  • sentence_good: a string feature.
  • sentence_bad: a string feature.
  • field: a string feature.
  • linguistics_term: a string feature.
  • UID: a string feature.
  • simple_LM_method: a bool feature.
  • one_prefix_method: a bool feature.
  • two_prefix_method: a bool feature.
  • lexically_identical: a bool feature.
  • pair_id: a int32 feature.

Data Splits

nametrain
adjunct_island1000
anaphor_gender_agreement1000
anaphor_number_agreement1000
animate_subject_passive1000
animate_subject_trans1000

Reference

We would like to acknowledge Warstadt, Alex and Parrish et al. for creating and maintaining the BLiMP dataset as a valuable resource for the computer vision and machine learning research community. For more information about the BLiMP dataset and its creator, please visit the BLiMP website.

License

The dataset has been released under the Creative Commons Attribution 4.0 (CC-BY-4.0) License.

Citation

@article{warstadt2020blimp,
    author = {Warstadt, Alex and Parrish, Alicia and Liu, Haokun and Mohananey, Anhad and Peng, Wei and Wang, Sheng-Fu and Bowman, Samuel R.},
    title = {BLiMP: The Benchmark of Linguistic Minimal Pairs for English},
    journal = {Transactions of the Association for Computational Linguistics},
    volume = {8},
    number = {},
    pages = {377-392},
    year = {2020},
    doi = {10.1162/tacl\_a\_00321},
    URL = {https://doi.org/10.1162/tacl_a_00321},
    eprint = {https://doi.org/10.1162/tacl_a_00321},
    abstract = { We introduce The Benchmark of Linguistic Minimal Pairs (BLiMP),1 a challenge set for evaluating the linguistic knowledge of language models (LMs) on major grammatical phenomena in English. BLiMP consists of 67 individual datasets, each containing 1,000 minimal pairs—that is, pairs of minimally different sentences that contrast in grammatical acceptability and isolate specific phenomenon in syntax, morphology, or semantics. We generate the data according to linguist-crafted grammar templates, and human aggregate agreement with the labels is 96.4\%. We evaluate n-gram, LSTM, and Transformer (GPT-2 and Transformer-XL) LMs by observing whether they assign a higher probability to the acceptable sentence in each minimal pair. We find that state-of-the-art models identify morphological contrasts related to agreement reliably, but they struggle with some subtle semantic and syntactic phenomena, such as negative polarity items and extraction islands. }
}