We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:


References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: BLiMP: The Benchmark of Linguistic Minimal Pairs for English

Abstract: We introduce The Benchmark of Linguistic Minimal Pairs (shortened to BLiMP), 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, and aggregate human agreement with the labels is 96.4%. We use it to evaluate n-gram, LSTM, and Transformer (GPT-2 and Transformer-XL) LMs. We find that state-of-the-art models identify morphological contrasts reliably, but they struggle with semantic restrictions on the distribution of quantifiers and negative polarity items and subtle syntactic phenomena such as extraction islands.
Comments: To appear in TACL
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:1912.00582 [cs.CL]
  (or arXiv:1912.00582v3 [cs.CL] for this version)

Submission history

From: Alex Warstadt [view email]
[v1] Mon, 2 Dec 2019 05:42:41 GMT (288kb,D)
[v2] Thu, 16 Apr 2020 02:07:03 GMT (642kb,D)
[v3] Wed, 23 Sep 2020 20:08:54 GMT (642kb,D)

Link back to: arXiv, form interface, contact.