Generalizable Negation Detection: Addressing the Challenges of Cross-Domain Data with Data Augmentation

Negation poses significant challenges for modern natural language processing (NLP) models, such as bidirectional encoder representations from transformers (BERT), often leading to misinterpretations. While the robust optimized BERT pre-training method (RoBERTa) is an optimized version of the BERT model, its ability to detect and interpret negations is limited, especially in complex linguistic and cross-domain environments. This work proposes a fine-tuning approach that utilizes counterfactual data augmentation (CDA) to improve the performance of the RoBERTa model in understanding and detecting negation. The proposed method incorporates both original and counterfactual samples, along with supplementary corpus generated by large language models (LLMs). By combining these diverse datasets, the approach improves the RoBERTa model's sensitivity to negation and its generalization across domains. Experimental results show that the developed approach taps the potential of counterfactual augmentation techniques to advance the NLP systems in addressing the challenges of negation.