A Comparative Evaluation of Thai Word Segmentation Techniques for Profanity Classification

Research aims to compare the performance of word segmentation techniques for Thai profanity classification. The dataset was collected from Social Media using an API, with a total of 3,538 tweets. These tweets were labeled manually and divided into two classes: profanity and non-profanity. The word segmentation techniques were PythaiNLP, ThaiTokenizer, OSKut, DeepCut, and ThaiLMCUT. PythaiNLP was the most accurate and correct when compared to the reference word. It had Precision, Recall, and F1-score values of 0.798, 0.838, and 0.818, respectively. Words obtained were then used to classify profanity using five algorithms: Logistic Regression (LR), Decision Trees (C4.5), K-nearest Neighbors (KNN), Support Vector Machine (SVM) and Long Short-Term Memory network (LSTM). The results showed that the words obtained from the PythaiNLP technique resulted in the SVM algorithm having the best performance for profanity classification. It had Accuracy and F1-Score values of 80.49% and 0.822, respectively. DeepCut, on the other hand, was found to have the lowest performance in terms of both Accuracy and F1-Score for almost all algorithms.