IPA Natural Language Inference

A 126-item BIG-bench task that asks for entailment, contradiction or neutral on MultiNLI sentence pairs written in IPA.

Also known as: IPA NLI, international phonetic alphabet nli

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.
Categoryreasoning
SubcategoryBIG-bench MultiNLI pairs transliterated into IPA
Page statusunknown
Metricmultiple_choice_grade
Directionhigher_is_better
Unit%
Dataset size126
Dataset licenceOANC End User License
PublisherGoogle (BIG-bench collaboration)

What it measures

international_phonetic_alphabet_nli presents two English sentences that have been converted to International Phonetic Alphabet transcriptions and asks whether the pair is entailment, contradiction or neutral. The authors, a Stanford student group, built it so that word-by-word IPA lookup is not enough: the model has to recover meaning in a rare symbol set. Source pairs are MultiNLI items restricted to author-generated hypotheses and OANC-licensed genres (fiction excluded). Transliteration uses the CMU Pronouncing Dictionary via the eng-to-ipa library.

Task format

Three-way multiple choice over entailment, contradiction and neutral. task.json sets preferred_score to multiple_choice_grade, append_choices_to_input false, and a task_prefix that defines the three labels. Canary GUID embedded. Keywords include zero-shot.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub