Negation plays a fundamental role in human communication and logical reasoning, yet it remains underrepresented in natural language inference (NLI) datasets. This work investigates the impact of targeted data augmentation using negation cues on the main NLI datasets for Portuguese (InferBR, ASSIN and ASSIN2). By synthetically generating new instances with negated hypotheses, we create more diverse training and test sets. A BERT-based model was fine-tuned and tested on the combined datasets and augmented data. The results show that the model was heavily influenced by the bias in the use of negation, and increased data diversity improves the model’s handling of negation.

Recent studies have questioned the ability of Large Language Models (LLMs) to handle logical negation. We revisit this issue within the Natural Language Inference (NLI) task, specifically investigating whether modern LLMs can distinguish negations that alter logical entailment (“important”) from those that do not (“unimportant”). For this purpose, we introduce NegNLI-BR, a new benchmark dataset in Portuguese designed to exercise this distinction. We evaluate a range of recent open-source LLMs, comparing the performance of their base and post-trained versions. Furthermore, we employ a causal probe to measure the Average Treatment Effect of negation interventions on the internal representations of LLMs. Our findings show that many recent LLMs, including smaller variants, effectively handle negation. The causal analysis reveals that important negations induce a stable and significant effect on model representations, distinct from unimportant negations or neutral filler words. We also observe that post-training generally enhances this representational sensitivity, suggesting it refines the models’ ability to encode the logical impact of negation.