INDONESIAN TARGETED ASPECT-BASED SENTIMENT ANALYSIS IN AUTOMOTIVE DOMAIN USING FEATURE MODIFICATION ON BI-LSTM

• Main Author: Ilmania, Arfinda
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39885
• Institution: Institut Teknologi Bandung

Targeted aspect-based sentiment analysis is a study which combines aspect and entity level sentiment analysis. In this task, both entitiy and aspect can appear more than one in a document. It makes this task able to provide more detailed analysis results. In this final project, a 2-step classification was developed to complete the task of the targeted aspect-based sentiment analysis system, namely aspect categorization and aspect sentiment classification. Modules are built by utilizing the Bi-LSTM topology and several position features such as positional embedding and context representation. Positional embedding and context representation are useful for providing additional information in the form of a pattern of the existence of an entity so that the module is easier to distinguish information about each entity. In addition, the neural pooling layer is also used to assist in the automatic extraction and selection of features. The other features that are also tried to support the performance of this system are POS tags and entity masking. Based on experiments that have been done, position features, neural pooling layer and POS tags are proven to improve the performance of the modules in this final project. Experiments carried out on the Indonesian language in the automotive domain. With 900 training data and 180 test data, the experimental results for the best module were obtained F1 score of 0.8838 by using positional embedding and POS tags on aspect categorization and context representation and POS tags on aspect sentiment classification. The results obtained can outperform the self-training baseline of Saeidi et al. (2016) who were retrained with the same dataset with the F1 score of 0.7779.