NeutralABSA: Enhancing Neutral Sentiment Classification in Aspect-Based Sentiment Analysis

Despite the widespread use of Aspect-Based Sentiment Analysis (ABSA) to extract fine-grained sentiments from text, a key problem that persists is the misclassification of neutral sentiments. Neutral sentiment represents the absence of a positive or negative evaluation towards an aspect, which may occur even when the text contains contextual or descriptive information. Unlike explicitly expressed sentiment, implicit sentiment may convey an underlying positive or negative evaluation without direct evaluative language. Existing ABSA models often struggle to distinguish genuinely neutral instances from those conveying implicit positive or negative sentiment, leading to incorrect polarity assignments. This study focuses on improving the classification of neutral sentiments by addressing existing challenges such as dataset imbalances, context overgeneralization and unrelated aspect-sentence pairs by introducing NeutralABSA. The datasets used for this research comprise the SemEval-2015 Task 12 Laptop and Restaurant domains as well as the FABSA dataset. The two main tasks addressed are Aspect Identification (AI) and Aspect Polarity Classification (APC). Depending on the granularity of dataset annotations, AI is performed either through explicit Aspect Term Extraction (ATE) or through classification into predefined aspect categories called as Aspect Category Classification (ACC). LoRA-based Parameter-Efficient Fine-Tuning (PEFT) is used to fine-tune BERT models to optimize AI and minimize context overgeneralization. Additionally, the issue of unrelated aspect-sentiment pairs was addressed by using data augmentation, improving the alignment between aspects and their corresponding sentiments. Neutral instances were assigned more weight during class-weighted DeBERTa training, reducing dataset imbalances and increasing the model’s ability to correctly categorize neutral sentiments. The proposed approach demonstrated improved precision in classifying neutral sentiments, achieving an overall 95% accuracy in Aspect Polarity Classification on the FABSA dataset.