Instagram online shop's comment classification using statistical approach

Abstract

Instagram is one of the currently popular social medias in Indonesia. Some online shop owners use Instagram to show their products and also use its comment section to communicate with their customers. A system that can classify Instagram comment automatically based on its response surely will help online shop owners to respond all various comments without having to read all of the comments. Thus, this study was conducted to find the best method to classify Instagram comment using statistical approach. In general, the system created in this study is divided into 4 components namely pre-process, feature extraction, feature selection, and classification. We compared several techniques in each component. In feature extraction, we compared lexical approaches (unigram) and word embedding approaches. As for the learning algorithm, we compared support vector machine (SVM) and convolutional neural network (CNN). The effect of pre-process and feature selection is also investigated in this study. The preprocess done in this study is basic pre-process, mention conversion, punctuation conversion, emoticon conversion, number conversion, formalization and region name conversion. The data used in this study is 2810 Instagram comments which have been labelled with 3 kinds of responses namely “answered”, “read”, and “ignored”. Using 10-fold cross validation, we conducted 3 experiment types, such as the baseline, pre-process, and word embedding. Baseline was done by using SVM as learning algorithm and unigram as feature. Pre-process experiment was done by adding pre-processing and feature selection to baseline. Word embedding experiment was done by using word embedding as feature and SVM or CNN as learning algorithm. The best result in this study was obtained from word embedding experiment using word embedding as feature representation, CNN as learning algorithm, and pre-processed model data with accuracy of 84.23%.