Advancing Aspect-Based Sentiment Analysis through Enhanced Contrastive Learning Techniques

Aspect-Based Sentiment Analysis (ABSA) is a method used to find out how people feel about specific parts (aspects) of something like the "features" of a laptop or the "service" at a restaurant within a sentence. It plays a big role in analyzing opinions in reviews. Recently, contrastive learning has become popular in improving ABSA. This learning method helps the system learn better by comparing examples like learning to tell the difference between a good and a bad review more clearly. This paper looks at two common contrastive learning methods for ABSA: Sentiment-Based Supervised Contrastive Learning: This method uses the actual sentiment labels (like "positive" or "negative" or “neutral”) to teach the model what to focus on. Augmentation-Based Unsupervised Contrastive Learning: This method creates new versions of the same sentence to help the model understand the meaning, without using sentiment labels.The paper also introduces four new methods to make ABSA even better: Prompt-Based Contrastive Learning (PromptCL): Uses AI models to create paraphrased sentences with the same meaning, helping the system learn from different ways of saying the same thing. Aspect-Specific Adversarial Contrastive Learning (ASACL): Slightly changes words near the aspect being analyzed, so the model becomes better at handling confusing or noisy inputs. Hierarchical Contrastive Learning (HiCL): Looks at both the whole sentence and specific parts to learn more complete understanding. Graph-Augmented Contrastive Learning (GraphCL): Uses graphs that show relationships between words to better understand how opinions are connected to aspects.