Convolutional Neural Networks for Text Classification: A Study on Public Activity Restriction
H Anggit Taba, Hari Suparwito(*)

Informatics Department
Universitas Sanata Dharma
Jl. Mrican Baru Caturtunggal Depok Sleman Yogyakarta

*shirsj[at]jesuits.net

Abstract

The COVID-19 pandemic two years ago has changed a lot in human life. To reduce the spread of the COVID-19 virus, the Indonesia government issued the Public Activity Restriction (PPKM). However, implementing PPKM caused polemics, pros and cons, in the community and was expressed through social media Twitter. The PPKM topic has become a trending topic on Twitter. There are so many opinions that it is difficult to classify which opinions are pros and cons with PPKM. This study aims to determine the pros and cons of PPKM through a Machine Learning approach using the Convolutional Neural Network (CNN) algorithm. The opinion data comes from Twitter tweets with restrictions on the keywords ppkm and ppkm rules. Data from Twitter have been collected from November 2021 to February 2022. We crawled tweets using tweepy, a python package to access the Twitter API and obtained 68,953 data. We deliberately choose data labelling without human intervention, so we label the tweet data using Vader, a python library, to determine whether the tweets have a negative, positive or neutral connotation. We proposed using the Wikipedia training Fasttext model to deliver word embedding. Next, the crucial step is to train datasets to get the predictive model. The implementation of the CNN algorithm focuses on layer architecture and parameter tuning variations. Accuracy results of 88% were obtained by using two convolution layers, ReLU and Softmax. Two Pooling techniques, MaxPooling and AveragePolling, were used to reduce the matrix size. It shows how to use machine learning approaches for predicting qualitative data in text processing.

Keywords: Convolutional Neural Network, Machine Learning, PPKM, Text Classification, Twitter

Topic: Computer Science