Aspect sentiment learning for Aspect-Level Sentiment Classification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence

Background:

• Aspect-Level Sentiment Classification (ALSC) is a key task in Sentiment Analysis (SA).
• Current methods often analyze sentences in isolation, missing crucial inter-sentence relationships for aspect sentiment.
• This limitation hinders a comprehensive understanding of aspect sentiment semantics.

Purpose of the Study:

• To introduce AspLearn, a novel aspect-learning method for ALSC.
• To optimize aspect sentiment semantics and generate robust aspect-specific sentence features.
• To enhance the performance of ALSC models by leveraging inter-sentence relationships.

Main Methods:

• AspLearn utilizes Aspect-aware Contrastive Learning (AspCL).
• AspCL mines aspect-related knowledge from relevant samples to refine aspect sentiment semantics.
• The method integrates this learned knowledge to improve sentence feature generation for ALSC.

Main Results:

• AspLearn demonstrated superior aspect learning capabilities across three benchmark datasets.
• The method achieved significant Macro F1 score improvements over existing state-of-the-art results on Laptops, Restaurants, and Twitter datasets.
• Experiments showed notable performance gains using DeBERTa as the backbone model.

Conclusions:

• AspLearn effectively optimizes aspect sentiment semantics and enhances ALSC performance.
• The method's ability to learn from inter-sentence relationships provides more robust aspect-specific features.
• AspLearn also shows potential in improving Large Language Models' (LLMs) sentiment recognition through relevant demonstration retrieval.