Empirical Study of Text Augmentation on Social Media Text in Vietnamese

Empirical Study of Text Augmentation on Social Media Text in Vietnamese Son T. Luu University of Information Technology VNU-HCM, Vietnam sonlt@uit.edu.vn Kiet Van Nguyen University of Information Technology VNU-HCM, Vietnam kietnv@uit.edu.vn Ngan Luu-Thuy Nguyen University of Information Technology VNU-HCM, Vietnam ngannlt@uit.edu.vn Abstract In the text classification problem, the im- balance of labels in datasets affect the per- formance of the text-classification models. Practically, the data about user comments on social networking sites not altogether appeared - the administrators often only allow positive comments and hide negative comments. Thus, when collecting the data about user comments on the social net- work, the data is usually skewed about one label, which leads the dataset to become imbalanced and deteriorate the model’s ability. The data augmentation techniques are applied to solve the imbalance problem between classes of the dataset, increasing the prediction model’s accuracy. In this paper, we performed augmentation tech- niques on the VLSP2019 Hate Speech De- tection on Vietnamese social texts and the UIT - VSFC: Vietnamese Students’ Feed- back Corpus for Sentiment Analysis. The result of augmentation increases by about 1.5% in the F1-macro score on both cor- pora. 1 Introduction In recent years, the growth of hate speech has become a crime, not only face-to-face action but also online communication (Fortuna and Nunes, 2018). The development of social net- work nowadays had made this situation worse. The threading of harassment comments and ha- rassment speech makes the user stop express- ing their opinions and looking up for other ideas (Vu et al., 2019). Fortuna and Nunes (2018) de- fined hate speech as the language attacking, di- minishing, and inciting violence or hate against individuals or groups based on their character- istics, religion, nations, and genders. To solve this problem, many datasets are constructed to detect and classify user comments on social net- work sites such as Facebook, Twitter, and Face- book in many languages1. The HSD-VLSP dataset (Vu et al., 2019) pro- vided by the VLSP 2019 Shared task about Hate speech detection on Social network2 con- tained nearly 25,000 comments and posts of Vietnamese Facebook users and has three labels. However, the distribution of three classes in the dataset is imbalanced. Besides, the UIT-VSFC dataset (Nguyen et al., 2018) that was used for predicting the feedback from students contained about 16,000 sentences and was annotated for two different tasks: sentiment analysis and topic analysis. Same as the HSD-VLSP dataset, the distribution of labels on the UIT-VSFC dataset is also imbalanced. We use the data augmen- tation techniques to generate new comments that belong to minority classes from the original dataset to tackle those restrictions. We conduct experiments on the augmented dataset and com- pare it with the original dataset to indicate data augmentation effectiveness. Those augmenta- tion techniques include synonym replacement, random insertion, random swapping, and ran- dom deletion (Wei and Zou, 2019). 1http://hatespeechdata.com/ 2https://www.aivivn.com/contests/8 The rests of the paper are structures as be- low. Section 2 introduces recent works in hate speech detection. Section 3 gives an …

Content truncated to 500 words