Not all Negation Cues are Equal: Affixal Negations Yield Better Negation Understanding

Details

Negation is a known weak point for language models, but prior research concentrated on a narrow set of high-frequency single-word cues like 'not' and 'never'. This leaves a large class of negation forms understudied: affixal negations embedded in word morphology (un-, non-, dis-, in-) and multi-word negation phrases. The authors build NegCue, a dataset with over 1.8 million samples spanning single-word, multi-word, and affixal negation across more than 200 unique cues. They pre-train both encoder-only language models and decoder-based LLMs on NegCue and evaluate on five downstream benchmarks. The key finding: negation types contribute unevenly to performance gains. Affixal negation yields the largest improvements. Single-word negation — the type most heavily studied — shows only modest gains despite being the focus of prior work. Pre-training on the full NegCue mix improves negation understanding for both model types. The paper is accepted to EMNLP 2026 main conference. The main contribution is the dataset itself: NegCue fills a genuine gap for researchers trying to evaluate or improve negation handling in LLMs beyond the basic 'not/never' case.