#TITLE_ALTERNATIVE#

• Main Author: JANOAH HASUDUNGAN - NIM:13514089, RAMOS
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/30197
• Institution: Institut Teknologi Bandung

<p align="justify">Researches on theclassification of emotions on onlinetext chat has not been done in Indonesian. The challenge of classifying emotions on online chatting text is the difference of characteristics on online chatting text with general text, such as unclear segmentation in the text, the existence of inter-document contexts affecting each other's labels, abnormal and multilingual vocabulary, and unclear syntacticstructures. To handle these matters, various adaptations of the text classification process in general, such as pre-processing of word normalization, pragmatic feature extraction, and not performing syntactic feature extraction are done in this final project. In addition, to handle the context of the sentence, in this final project, there aretwo approaches, non-sequential approach and sequential approach. In the non-sequential approach, the context between documents will not be taken into account, whereas in the sequential approach, the context between chat bubbles will be taken into account, whereas in the sequential approach, the context betweenchatbubbleswillhaveaneffect. In both of these approaches, experiments were conducted on the combination of features and machine learning algorithms. The features are a combination of ngram, pragmatic features, non-textual features, and word embedding features. The learning algorithms tried are K-nearest neighbor (KNN), support vector machine (SVM), multilayer perceptron (MLP), and random forest for non-sequential approachalsoLongShort-TermMemory(LSTM)forsequentialapproach. Model building and testing were performed using data derived from social media LINE, by splitting the data into training, validation, and testing data. The best F1micro emotion evaluated on the test and validation data is 0.326 and 0.325 respectively using multilayer perceptron with 1-gram, pragmatic feature and averagewordvectorwordembeddingfeature.<p align="justify"> <br />