Graph-enhanced implicit aspect-level sentiment analysis based on multi-prompt fusion.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Implicit Aspect-Level Sentiment Analysis (IALSA) identifies unstated opinions in text.
• Challenges in IALSA include omitted expressions and limited features in short texts.
• Existing methods struggle with semantic comprehension of implicit sentiment.

Purpose of the Study:

• To propose a novel generative model for graph-enhanced IALSA.
• To improve the accuracy and robustness of identifying implicit aspects and opinions.
• To address the challenges posed by omitted expressions and short texts in sentiment analysis.

Main Methods:

• Utilized a generative pre-training model (T5) combined with graph neural networks.
• Employed multi-prompt fusion to leverage complementary information from various prompts.
• Designed aspect and opinion term recognition as index generation tasks with a differentiated loss function.

Main Results:

• Achieved a 1.99% F1 score improvement on the restaurantACOS dataset.
• Obtained an 1.83% F1 score improvement on the LaptopACOS dataset.
• Demonstrated effectiveness of individual improvements through ablation experiments, showing benefits in low-resource settings.

Conclusions:

• The proposed graph-enhanced generative model significantly advances IALSA.
• Multi-prompt fusion and specialized task design enhance the understanding of implicit sentiment.
• The model shows strong performance, especially in low-resource scenarios, offering a robust solution for complex sentiment analysis.