LM-SynEval (Targeted Syntactic Evaluation of Language Models)

72 auto-generated minimal-pair test sets probing whether a model's probabilities favour the grammatical sentence for subject-verb agreement, reflexive anaphora and negative polarity items.

Also known as: Targeted Syntactic Evaluation of Language Models, Marvin and Linzen 2018

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryknowledge
Subcategorytargeted syntactic minimal pairs (agreement, reflexives, negative polarity items)
Page statusactive
Metricpairwise accuracy (grammatical sentence assigned the higher probability)
Directionhigher_is_better
Unit%
Dataset size158084
Dataset licenceMIT

What it measures

LM-SynEval measures what a language model implicitly knows about specific points of English syntax, not whether it can perform a task or explain a rule. Each item is a minimal pair of sentences differing in one grammatical property -- for example "The author laughs" versus "The author laugh" -- and the model is scored by whether it assigns a higher probability to the grammatical member of the pair. The pairs are organised into three phenomena: subject-verb agreement (tested across many constructions -- across a prepositional phrase, a sentential complement, subject and object relative clauses with and without an overt "that", and verb-phrase coordination), reflexive anaphora (does the reflexive pronoun's number match its antecedent, again tested within simple sentences and across relative clauses), and negative polarity item licensing (does "ever" or a similar NPI appear only where a licensing context such as "no" makes it grammatical). A model can score well on every construction here while being unable to state, in words, the agreement or licensing rule it is implicitly satisfying -- this is a probe of linguistic competence, closer to a psycholinguistic acceptability experiment than to a benchmark of task-solving ability like question answering.

Task format

Forced-choice by probability comparison: for each minimal pair, compare the model's log-probability on the grammatical sentence against its ungrammatical, minimally different counterpart; the harness presents exactly two choices per item and no explicit answer or generation is requested.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub