Chi-Square Feature Selection with Pseudo-Labelling in Natural Language Processing

Abstract

This study aims to evaluate the effectiveness of the Chi-Square feature selection method in improving the classification accuracy of linear Support Vector Machine, K-Nearest Neighbors and Random Forest in natural language processing when combined with classification algorithms as well as introducing Pseudo-Labelling techniques to improve semi-supervised classification performance. This research is important in the context of NLP as accurate feature selection can significantly improve model performance by reducing data noise and focusing on the most relevant information, while Pseudo-Labelling techniques help maximise unlabelled data, which is particularly useful when labelled data is sparse. The research methodology involves collecting relevant datasets, thus applying the Chi-Square method to filter out significant features, and applying Pseudo-Labelling techniques to train semi-supervised models. In this study, the dataset used in this research is the text data of public comments related to the 2024 Presidential General Election, which is obtained from the Twitter scrapping process. The characteristics of this dataset include various comments and opinions from the public related to presidential candidates, including political views, support, and criticism of these candidates. The experimental results show a significant improvement in classification accuracy to 0.9200, with precision of 0.8893, recall of 0.9200, and F1-score of 0.8828. The integration of Pseudo-Labelling techniques prominently improves the performance of semi-supervised classification, suggesting that the combination of Chi-Square and Pseudo-Labelling methods can improve classification systems in various natural language processing applications. This opens up opportunities to develop more efficient methodologies in improving classification accuracy and effectiveness in natural language processing tasks, especially in the domains of linear Support Vector Machine, K-Nearest Neighbors and Random Forest well as semi-supervised learning.

Keywords

• chi-square feature selection
• natural language pocessing
• pseudo-labelling
• semi-supervised.