Search for collections on Undip Repository

ANALISIS SENTIMEN BERBASIS ASPEK UNTUK ULASAN PENGGUNA MARKETPLACE MENGGUNAKAN STANZA DAN PENDEKATAN SEMI-SUPERVISED LEARNING

CHAMID, Ahmad Abdul and Widowati, Widowati and Kusumaningrum, Retno (2024) ANALISIS SENTIMEN BERBASIS ASPEK UNTUK ULASAN PENGGUNA MARKETPLACE MENGGUNAKAN STANZA DAN PENDEKATAN SEMI-SUPERVISED LEARNING. Doctoral thesis, UNIVERSITAS DIPONEGORO.

[img] Text
1-COVER_compressed.pdf

Download (8MB)
[img] Text
2-BAB I_compressed.pdf

Download (69kB)
[img] Text
3-BAB II_compressed.pdf

Download (855kB)
[img] Text
4-BAB III_compressed.pdf
Restricted to Repository staff only

Download (798kB)
[img] Text
5-BAB IV_compressed.pdf
Restricted to Repository staff only

Download (741kB)
[img] Text
6-BAB V_compressed.pdf
Restricted to Repository staff only

Download (58kB)
[img] Text
7-DAFTAR PUSTAKA_compressed.pdf

Download (124kB)
[img] Text
8-LAMPIRAN_compressed.pdf
Restricted to Repository staff only

Download (1MB)

Abstract

Ulasan yang ditampilkan sebagai bintang di marketplace dianggap kurang informatif dan dapat menimbulkan disinformasi. Oleh karena itu diperlukan suatu informasi yang andal yang dapat diwujudkan melalui Aspect-Based Sentiment Analysis (ABSA). ABSA bertugas mendeteksi aspek dan mengklasifikasikan sentimen. Penelitian ABSA di Indonesia telah berkembang dengan baik menggunakan classical machine learning dan deep learning. Pendekatan classical machine learning menghadapi kendala dimana keberhasilan deteksi aspek sangat bergantung pada kelengkapan leksikon yang dibangun. Oleh karena itu, penelitian ABSA selanjutnya berfokus pada penerapan deep learning. Namun penerapan deep learning masih perlu ditingkatkan dengan adanya inkonsistensi deteksi aspek. Dalam penelitian ini, pendekatan dependency parser diusulkan untuk mengatasi inkonsistensi deteksi aspek dengan memanfaatkan kemampuan stanza dalam menguraikan struktur tata bahasa. Keberhasilan deteksi kata benda dianggap penting untuk deteksi aspek. Pendekatan deep learning telah digunakan untuk klasifikasi sentimen, yang terbukti kuat, namun kendala utamanya adalah terbatasnya data berlabel, sehingga sulit diperoleh. Solusinya adalah dengan menggunakan pendekatan semi-supervised yang memanfaatkan sejumlah kecil data berlabel dan sejumlah besar data tidak berlabel. Data ulasan produk kaos dari marketplace di Indonesia yang digunakan dalam penelitian mencapai 15.237 data ulasan. Proses selanjutnya meliputi pra-pemrosesan data, pembagian data, dan penguraian struktur tata bahasa kalimat ulasan menggunakan stanza. Dua ahli bahasa secara manual memberi label pada sebagian kecil data (28%) dengan hasil konsistensi yang sangat baik. Data berlabel digunakan untuk melatih model deteksi aspek dan klasifikasi sentimen menggunakan metode Neural Network (NN), Convolutional Neural Network (CNN), dan Recurrent Neural Network (RNN). Berdasarkan percobaan, model terbaik untuk deteksi aspek menggunakan metode RNN dengan F1-score sebesar 0,980, sedangkan model terbaik untuk klasifikasi sentimen menggunakan metode CNN dengan F1-score sebesar 0,876. Hasilnya, sistem atau aplikasi analisis sentimen berbasis aspek telah berhasil dibangun dengan menggunakan model terbaik yaitu ABSA v2.0. Sistem ini dapat dimanfaatkan oleh pihak terkait untuk lebih memahami sentimen yang terkait dengan aspek tertentu dari ulasan produk.
Kata kunci: sentimen analisis berbasis aspek, semi-supervised, deep learning, dan stanza.

Reviews displayed as stars on the marketplace are considered less informative and can give rise to disinformation. Therefore, a reliable system is needed, which can be realized through Aspect-Based Sentiment Analysis (ABSA). ABSA is tasked with aspect detection and sentiment classification. ABSA research in Indonesian has developed well using classical machine learning and deep learning. The classical machine learning approach faces obstacles where the success of aspect detection is very dependent on the completeness of the built lexicon. Therefore, ABSA's next research focuses on the application of deep learning. However, the application of deep learning still needs to improve with aspect detection inconsistencies. In this research, a dependency parser approach is proposed to overcome aspect detection inconsistencies by utilizing the ability of stanzas to decipher grammatical structures. Successful detection of nouns is considered crucial for aspect detection. A deep learning approach has been used for sentiment classification, which has proven robust, but the main obstacle is the limited labeled data, which is difficult to obtain. The solution is to use a semi-supervised approach utilizing a small amount of labeled data and a large amount of unlabeled data. Tshirt product review data from marketplaces in Indonesia was used in the research, reaching 15,237 review data. The following process involves pre-processing the data, dividing the data, and deciphering the grammatical structure of the review sentences using stanzas. Two linguists manually labeled a small portion of the data (28%) with excellent consistency results. Labeled data is used to train aspect detection and sentiment classification models using Neural Network (NN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) methods. Based on experiments, the best model for aspect detection uses the RNN method with an F1-score of 0,980, while the best model for sentiment classification uses the CNN method with an F1-score of 0,876. As a final result, an aspect-based sentiment analysis system has been successfully built using the best model, namely ABSA v2.0. This system can be leveraged by interested parties to understand better the sentiment associated with certain aspects of product reviews.
Keywords: aspect-based sentiment analysis, semi-supervised, deep learning, and stanza.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: sentimen analisis berbasis aspek, semi-supervised, deep learning, dan stanza.
Subjects: Sciences and Mathemathic
Divisions: Postgraduate Program > Doctor Program in Information System
Depositing User: ekana listianawati
Date Deposited: 13 May 2024 02:37
Last Modified: 13 May 2024 02:37
URI: https://eprints2.undip.ac.id/id/eprint/22851

Actions (login required)

View Item View Item