Not All Negation Cues Are Equal: Improving Negation Understanding by Leveraging All Cue Types

ACL ARR 2026 January Submission9149 Authors

Keywords: Negation, Negation Cues, Natural Language Understanding, Language Models, Pre-training, Semantics

Abstract: This paper investigates whether pre-training with affixal, single-word, and multi-word negation cues results in improved negation understanding. We work with encoder-only LMs as well as smaller LLMs. Experimental results show that pre-training with affixal negations is most beneficial across all negation types at inference time, and benefits persist regardless of cue frequency. Our pre-training is model- and task-agnostic; we show improved results across all models and five benchmarks for question answering, information retrieval and inference.

Paper Type: Long

Research Area: Language Models

Research Area Keywords: pre-training, robustness, transfer, fine-tuning, prompting

Contribution Types: NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 9149