Quantum-Inspired Complex-Valued Language Models for Aspect-Based Sentiment Classification.

Area of Science:

• Natural Language Processing
• Quantum Computing Applications

Background:

• Aspect-based sentiment analysis (ABSA) is a fine-grained task focused on classifying sentiment polarities for specific aspects.
• Current ABSA models predominantly use real-valued vector representations, potentially limiting their capacity to capture nuanced information.

Purpose of the Study:

• To investigate the efficacy of Hilbert Space representations for ABSA models.
• To explore the complexification of existing real-valued models to enhance sentiment analysis capabilities.

Main Methods:

• The study involves constructing complex-valued versions of three strong real-valued baseline models.
• Mathematical formulations from quantum theory are utilized to represent words as complex-valued vectors in a Hilbert Space.

Main Results:

• Experimental results validate the effectiveness of complexification in ABSA.
• The proposed complex-valued models demonstrate superior performance compared to their real-valued counterparts.
• Complex-valued embeddings were found to convey additional information beyond real embeddings.

Conclusions:

• Complex-valued embeddings offer a promising avenue for advancing aspect-based sentiment analysis.
• The complex-valued RoBERTa model achieves state-of-the-art or competitive results on benchmark ABSA datasets.