Joint Distribution pada Weighted Majority Vote (WMV) untuk Peningkatan Kinerja Sentiment Analysis Tersupervisi pada Dataset Twitter

Abstrak

Sentiment analysis adalah teknik komputasi text mining berbasis natural language processing (NLP) untuk mengekstraksi pendapat seseorang yang diungkapkan dalam platform online, termasuk dalam platform microblogging Twitter, salah satu platform microblogging yang paling popular digunakan di Indonesia. Ada dua pendekatan yang umum digunakan dalam teknik sentiment analysis yaitu pendekatan berbasis machine learning (ML) dan pendekatan berbasis sentiment lexicon (SL). Fokus penelitian ini adalah untuk pengembangan teknik sentiment analysis berbasis machine learning yang disebut juga teknik tersupervisi pada dataset Twitter. Sebagian besar sentiment analysis pada dataset Twitter berbahasa Indonesia mengandalkan single machine learning algorithm. Penelitian ini menggabungkan kinerja berbagai algoritma/experts seraya mengurangi tingkat kesalahan klasifikasi dengan meng-update bobot secara dinamis menggunakan weighted majority vote (WMV) berbasis joint distribution dari Bayesian Network. Pada tahap pertama, data di grabbing dari Twitter dengan 3 hashtag terkait Covid-19 sebagai data eksperimen. Selanjutnya kinerja weighted majority vote secara ekstensif dibandingkan dengan 4 metode baseline sebagai pembanding, yaitu: Naïve Bayes, Gaussian Naïve Bayes, Multinomial Naïve Bayes dan Majority Vote dari ketiga single classifier tersebut. Metrics kinerja yang digunakan adalah precision, recall, fmeasure, accuracy dan Mathews correlation coeficient (MCCC). Dalam eksperimen, terbukti bahwa WMV mampu meningkatkan kinerja sentiment analysis pada ketiga topik dataset dengan evaluator berbagai metrics kinerja sentiment analysis.

Abstract

Sentiment analysis is a computational text mining technique based on natural language processing (NLP) to extract someone's opinion expressed in online platforms, including the Twitter microblogging platform, one of the most popular microblogging platforms used in Indonesia. There are two approaches that are commonly used in sentiment analysis techniques, namely the machine learning (ML) based approach and the sentiment lexicon (SL) based approach. The focus of this research is the development of machine learning-based sentiment analysis techniques which are also called supervised techniques on the Twitter dataset. Most of the sentiment analysis on the Indonesian language Twitter dataset relies on a single machine learning algorithm. This study combines the performance of various algorithms/experts while reducing the level of misclassification by updating the weights dynamically using a joint distribution-based weighted majority vote (WMV) from the Bayesian Network. In the first stage, data was grabbed from Twitter with 3 hashtags related to Covid-19 as experimental data. Furthermore, the performance of the weighted majority vote was extensively compared with 4 baseline methods for comparison, namely: Naïve Bayes, Gaussian Naïve Bayes, Multinomial Nave Bayes and Majority Vote from the three single classifiers. Performance metrics used are precision, recall, fmeasure, accuracy and Mathews correlation coeficient. In experiments, it is proven that WMV is able to improve sentiment analysis performance on the three dataset topics with various evaluators of sentiment analysis performance metrics.