Samuel R. Bowman · page 2
Papers
- Counterfactually-Augmented SNLI Training Data Does Not Yield Better Generalization Than Unaugmented Data (2020 · 51 citations)
Crowdsourced NLI benchmarks contain shortcuts such as hypothesis-only cues, negation words and lexical overlap. - More Constructions, More Genres: Extending Stanford Dependencies (2013 · 47 citations)
The Stanford scheme was built on newswire and gives little analysis to constructions that are rare there but common in web text: tough adjectives, free relatives, comparatives, small clauses, questions, imperatives, fragments and lists. - NOPE: A Corpus of Naturally-Occurring Presuppositions in English (2021 · 40 citations)
NOPE collects about 1,280 corpus passages, each with a presupposition trigger of one of ten types: change of state, aspectual verbs (begin, stop), clause-embedding verbs (know, realize, say), clefts, comparatives, embedded questions, implicatives (manage to, fail to), numeric determiners (both), re- verbs, and… - New Protocols and Negative Results for Textual Entailment Data Collection (2020 · 27 citations)
Data from natural language inference (NLI) is widely used for intermediate training before fine-tuning on other understanding tasks. - Two Failures of Self-Consistency in the Multi-Step Reasoning of LLMs (2023 · 25 citations)
Evaluations of LLMs usually check whether answers are correct, not whether a model's answers agree with each other. - BLiMP: The Benchmark of Linguistic Minimal Pairs for English (2020 · 21 citations)
BLiMP is a challenge set for measuring which parts of English grammar language models have learned. It contains 67 paradigms of 1,000 sentence pairs each; within a pair the two sentences differ minimally and only one is acceptable. The pairs are generated automatically from templates designed by linguists and cover…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/samuel-r-bowman/page/2/