Sentiment Classification of Chinese Tourism Reviews Based on ERNIE-Gram+GCN

Area of Science:

• Natural Language Processing
• Machine Learning
• Computer Science

Background:

• Tourists increasingly rely on online reviews, including emojis, for travel decisions.
• Accurate classification of attraction reviews is crucial for user experience and business insights.
• Existing models may not effectively capture the nuances of emoji-inclusive text data.

Purpose of the Study:

• To develop a high-accuracy model for classifying Chinese attraction reviews that incorporate emojis.
• To introduce the Chinese Attraction Evaluation Incorporating Emojis (CAEIE) dataset.
• To propose and evaluate the E2G model, combining ERNIE-Gram and Text Graph Convolutional Network (TextGCN).

Main Methods:

• Construction of the Chinese Attraction Evaluation Incorporating Emojis (CAEIE) dataset.
• Development of the E2G model, integrating ERNIE-Gram's masked language modeling with TextGCN's graph-based text representation.
• Utilizing an explicitly n-gram masking method to enhance information integration.
• Comparison with advanced models and evaluation of different activation functions for TextGCN.

Main Results:

• The E2G model achieved a classification accuracy of 97.37% on the CAEIE dataset.
• E2G outperformed ERNIE-Gram by 1.37% and TextGCN by 1.35% in accuracy.
• TextGCN demonstrated superior performance when combined with ERNIE-Gram (1.6% higher) and TextGAT (2.15% higher).
• Rectified linear unit 6 (RELU6) was identified as the optimal activation function for the second layer of TextGCN.

Conclusions:

• The E2G model effectively classifies Chinese attraction reviews containing emojis, demonstrating high accuracy.
• The proposed n-gram masking method enhances the integration of coarse-grained information in pre-training.
• The study highlights the importance of specialized datasets and models for analyzing user-generated content with rich linguistic features like emojis.