sEntIMeldCL: Enhancing explicit knowledge via Uniform-based Implicit Contrastive Mechanism for Aspect-Level Sentiment...

Aspect-level sentiment analysis (ALSA) is a fine-grained task which consists of aspect terms and sentiment polarities within sentences. Numerous research studies only focus on the syntactical dependencies between words, ignoring the impact of negations on sentiment polarity. Several pre-trained augmentation models play an essential role in solving this problem. However, those studies use word substitution to replace words with synonyms, often neglecting the semantic representation of multiple aspects within a sentence. Therefore, we propose a Uniform-based Implicit Contrastive Mechanism to enhance the explicit knowledge for ALSA called sEntIMeldCL-ALSA to overcome the abovementioned challenges. This work aims to extract the semantic correlations between aspect terms and sentence polarity, directly connected during data augmentation. Our proposed model comprises three modules to harness semantic knowledge for the ALSA task. The aspect-specific segmentation adapter module segments the sentence to extract a term-specific explicit representation. Subsequently, the uniform-based implicit augmentation module creates polarity-dependent augmented sentences and enhances the semantic representation of explicit data. A dual contrastive loss called ExImp contrastive module has been utilized to enhance the model's performance further. After obtaining explicit and implicit representations, we concatenate them to form a unified semantic representation of the data. Extensive results on four benchmark datasets indicated that the proposed sEntIMeldCL achieved state-of-the-art performance on three datasets, enhancing accuracy up to 87.37%, 81.25%, 77.62%, 85.25%, while the F1 score improved by 81.87%, 77.56%, 76.74%, 84.69% on Restaurant, Laptop, Twitter, and MAMS datasets, respectively.