Text Augmentation for Indonesian Intent Classification: Comparative Study

Several problems occur when doing intent classification, one of them is imbalanced data. Augmentation could be the solution to this problem. There are several methods of augmentation but there has not been any paper compared specific types of augmentation directly to know which method give the best result, especially when the data is in Indonesian language. Moreover, the Indonesian language used in the data is often non-formal. There has been no research that has compared augmentation techniques for Indonesian language. This article compares three types of augmentation: NLP augmentation, Large Language Models, and Easy Data Augmentation to determine whether augmentation affects the intent classification results or not in Indonesian language. This paper also uses a classifier known as Bidirectional Encoder Representations from Transformers (BERT). After several steps from data processing until implementing those 3 augmentation methods, the results show that there was an increase in the data before and after being augmented. LLM is the augmentation method with the best level of accuracy for agent intention with a value of 76.65% and NLPAUG is the best augmentation for visitor intention with a value of 72.26%.