Performance analysis of aspect-level sentiment classification task based on different deep learning models.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Aspect-level sentiment classification (ASCT) models typically assume uniform data distribution.
• Retraining models for new data distributions is costly and labor-intensive, especially for aspect-level annotation.
• Deep learning models show promise but face challenges with cross-lingual and distribution shift scenarios.

Purpose of the Study:

• To systematically analyze and compare the performance of various deep learning models for ASCT.
• To investigate model performance across different aspect quantities, computational costs, and specific cases.
• To propose and evaluate a novel ASCT method that facilitates cross-language migration and addresses data distribution challenges.

Main Methods:

• Comparative analysis of sequence-based, graph-based convolutional neural networks, and pre-training language models.
• Evaluation on eight public datasets in Chinese and English, assessing classification performance, aspect number impact, case studies, and computational cost.
• Design and implementation of a state-of-the-art ASCT classification method.

Main Results:

• Performance variations observed across different model architectures and datasets.
• The proposed method demonstrates potential for improved cross-language migration.
• Insights gained into model robustness concerning aspect numbers and computational efficiency.

Conclusions:

• Deep learning models offer effective solutions for ASCT, but cross-lingual transfer remains a challenge.
• The developed ASCT method provides a promising direction for overcoming data distribution shifts and enabling cross-lingual applications.
• Further research is needed to address model limitations and explore advanced migration strategies.