A simple interpolation-based data augmentation method for implicit sentiment identification.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Implicit sentiment identification is crucial in NLP but faces challenges like semantic complexity and data scarcity.
• Deep learning models struggle with implicit sentiment due to lack of explicit sentiment words and limited annotated data.
• Existing data augmentation methods often lack validity and can introduce noise.

Purpose of the Study:

• To propose an effective data augmentation method for implicit sentiment identification.
• To address the data scarcity and overfitting issues in implicit sentiment analysis.
• To improve the performance of models in identifying implicit sentiment.

Main Methods:

• Developed ISIMIX, an interpolation-based data augmentation technique operating in hidden space.
• Generated augmented samples through underlying augmentation techniques and blended them with original data.
• Incorporated Jensen-Shannon divergence regularization to minimize discrepancies between raw and augmented data distributions.

Main Results:

• ISIMIX effectively alleviates the data scarcity problem in implicit sentiment identification.
• The method demonstrates superior performance compared to existing data augmentation techniques.
• Experimental results on three public datasets show ISIMIX outperforms mainstream text classification methods.

Conclusions:

• ISIMIX is a simple yet effective approach for implicit sentiment identification.
• The method shows significant potential for wide application in NLP tasks.
• Interpolation in hidden space offers a valid and robust solution for data augmentation.