One for "All": a unified model for fine-grained sentiment analysis under three tasks.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Fine-grained sentiment analysis extracts consumer emotions towards specific entities and aspects from text.
• Existing tasks include Aspect-Based Sentiment Analysis (ABSA), Target-Based Sentiment Analysis (TABSA), and Multi-Entity Aspect-Based Sentiment Analysis (MEABSA).
• Common challenges in these tasks are data scarcity (low-resource problem) and skewed sentiment polarity.

Purpose of the Study:

• To propose a unified model addressing low-resource issues and sentiment polarity bias in fine-grained sentiment analysis.
• To develop a model applicable across ABSA, TABSA, and MEABSA tasks.

Main Methods:

• A unified model, PEA (Pre-trained language model with Ensemble and Augmentation), is proposed.
• Data augmentation techniques, including entity replacement and dual noise injection, are employed to tackle data scarcity and bias.
• An ensemble method combines results from RNN-based and BERT-based models.

Main Results:

• PEA demonstrates significant performance improvements across all three fine-grained sentiment analysis tasks compared to state-of-the-art models.
• The model achieves comparable results to baselines using only 20% of the training data.
• This highlights PEA's effectiveness in extreme low-resource conditions.

Conclusions:

• The proposed PEA model offers a robust solution for fine-grained sentiment analysis, particularly under low-resource constraints.
• PEA successfully mitigates sentiment polarity bias and enhances model performance.
• The unified approach provides a versatile tool for analyzing diverse social media data.