AF'IDAH, Dwi Intan and Kusumaningrum, Retno and Surarso, Bayu (2020) LONG SHORT TERM MEMORY - CONVOLUTIONAL NEURAL NETWORK PADA ANALISIS SENTIMEN ULASAN OBJEK WISATA DI PULAU BALI BERBAHASA INDONESIA. Masters thesis, School of Postgraduate Studies.

	Text FINAL TESIS_DWI INTAN A cover - abstract-pages-1-15.pdf Download (2MB)
	Text FINAL TESIS_DWI INTAN A bab 1 -pages-16-18.pdf Download (2MB)
	Text FINAL TESIS_DWI INTAN A bab 2 -pages-19-40.pdf Download (2MB)
	Text FINAL TESIS_DWI INTAN A bab 3 -pages-41-75.pdf Restricted to Repository staff only Download (2MB)
	Text FINAL TESIS_DWI INTAN A bab 4 -pages-76-96.pdf Restricted to Repository staff only Download (3MB)
	Text FINAL TESIS_DWI INTAN A bab 5 - lampiran -pages-97-190.pdf Restricted to Repository staff only Download (3MB)

Abstract

Jumlah data yang besar menjadi tantangan tersediri bagi proses analisis sentimen. Deep learning mampu mengatasi masalah analisis data dalam jumlah besar dengan memberikan performa yang lebih baik dibandingkan machine learning klasik. Model deep learning berupa Long Short Term Memory (LSTM) dan Convolutional Neural Network (CNN) masing-masing memiliki performa yang baik pada penelitian analisis sentimen. Akan tetapi, keduanya memiliki masalah yang perlu diselesaikan. LSTM memiliki kelemahan tidak mampu menangkap informasi yang dominan. Sementara CNN memiliki kelemahan tidak mampu memahami ketergantungan semantik jarak jauh. Masalah lain yang muncul dalam analisis sentimen adalah penentuan model pre-training yang tepat agar diperoleh model yang akurat. Oleh karena itu, penelitan ini mengusulkan metode LSTM-CNN untuk analisis sentimen dengan model pre-training menggunakan Word2Vec pada dokumen teks berbahasa Indonesia. Word2Vec dipilih karena dapat menangkap makna semantik teks dengan baik dan setiap kata yang berhubungan dicirikan dengan vektor yang cenderung mirip. Dataset yang digunakan pada penelitian ini berupa ulasan objek wisata di Pulau Bali berbahasa Indonesia yang berjumlah 10000 ulasan dengan 5000 ulasan positif dan 5000 ulasan negatif. Kombinasi parameter dari LSTM-CNN dan Word2Vec yang diteliti antara lain pooling layer, dropout, learning rate, aktivasi konvolusi, arsitektur Word2Vec, metode evaluasi Word2Vec, dan dimensi Word2Vec. Rata-rata akurasi dari model terbaik LSTM-CNN dan Word2Vec sebesar 97.17% yang diperoleh dari kombinasi parameter LSTM-CNN berupa averagepool sebagai pooling layer, droput 0,7, learning rate 0,0001, ReLU sebagai konvolusi activasi, dan kombinasi parameter Word2Vec berupa Skip-gram, Hierarchical Softmax, dan dimensi 300.
Kata Kunci: Long Short Term Memory, Convolutional Neural Network, Deep Learning, Analisis Sentimen, Ulasan Objek Wisata

Huge amount of text present a challenge for the sentiment analysis process. Deep learning is able to solve this challenge by providing better performance in the sentimen analysis than classic machine learning which uses statistical method. Long Short Term Memory (LSTM) and Convolutional Neural Network (CNN) as deep learning have shown well performance for the sentimen analysis study. However, both have problems which need to be resolved. LSTM has inability to capture diminan information. While CNN has weakness to understand semantic dependencies. Another problem in sentiment analysis is the determination of the right pre-training model in order to obtain an accurate model. Therefore, this research proposes the LSTM-CNN method for sentiment analysis with a pre-training model using Word2Vec in Indonesian text document. Word2Vec was chosen since it can capture the semantic meaning text well and each related word is characterized by vectors that tend to be similar. The dataset is used in the study is 10000 reviews of Bali’s touristic destination in Indonesian language which comprise 5000 positive review and 5000 negative review. The combination of LSTM-CNN and Word2Vec parameters that were tested in the research are pooling layer, dropout, learning level, convolutional activation, architecture of Word2Vec, evaluation method of Word2Vec, and dimension of Word2Vec. The result of this research show that accuracy of the best LSTM-CNN and Word2Vec model is 97.17% which obtained by using LSTM-CNN parameter namely averagepool as pooling layer, dropout of 0,7, learning rate of 0,0001, ReLU as convolutional activation, and using Word2Vec parameter namely Skip-gram, Hierarchical Softmax, dimension of 300.
Keywords: Sentiment Analysis, Long Short Term Memmory, Convolutional Neural Network, Word2Vec, Touristic Destination Review

Actions (login required)

View Item