ASPECT-BASED SENTIMENT ANALYSIS FOR GAME REVIEW USING WEAKLY SUPERVISED LEARNING

• Main Author: Muhamad Gana, Hanif
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78294
• Institution: Institut Teknologi Bandung

Each player gives different importance to different aspects of different genres, and finding which game to play that has the important aspects is usually done reading game reviews. Aspect-based sentiment analysis (ABSA) can be employed on the reviews to simplify this search, however resources to do so for the domain is scarce, especially for Indonesian game reviews. One alternative would be using ABSA methods that employ weakly supervised or unsupervised learning. One ABSA method for English reviews that employ weakly supervised learning is Joint-Aspect Sentiment Topic Embedding (JASen) (Huang et al., 2020), that only requires topic list and their keywords and a Word2Vec-based (Mikolov et al., 2013) pretrained embeddings for training, and can create a joint model where aspect and sentiment can be predicted at once. This paper aims to adapt JASen for Indonesian game reviews by changing its preprocessing, and reimplementing and modifying its training process to increase compatibility with Indonesian texts. From experiments performed it is found that the best configuration was possibly one that has kernel heights of [1, 3, 4], used a hybrid-domain pretrained embedding, is split into separate aspect and sentiment pair of models, and trained specifically on its genre. The macro-F1 score of said configuration to the test data for aspect categorization is 56.55, while the score for sentiment classification is 56.34.