TERMS AND POLARITY CO-EXTRACTION FOR ASPECT BASED SENTIMENT ANALYSIS

• Main Author: Ahmad Genadi, Rifo
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/48329
• Institution: Institut Teknologi Bandung

Aspect-based sentiment analysis (ASBA) can help companies and the general public to find out opinions about a product. Aspect-based sentiment analysis can be done with a variety of approaches, and the tasks that are carried out are diverse, consisting of aspect term extraction, aspect categorization, sentiment term extraction, sentiment polarity classification, and extraction of aspect and sentiment term relations. The tasks performed are generally carried out one by one in sequence. Such method is considered less efficient and can reduce the performance of the model due to errors in previous processes. Co-extraction approach with DOER architecture achieve better performance than pipelined approach. This Final Project focuses on adapting the coextraction approach and modifying it so that the extraction of aspect terms, sentiment terms, and sentiment polarity is done simultaneously. The co-extraction approach was adapted by modifying the DOER architecture to perform the sentiment expression expression task also, namely by modifying the output layer and train the adapted model on a collection of Indonesian-language hotel reviews. The adaptation is done by testing the output layer topology for aspect and sentiment term extraction as well as variations in the type of RNN cells used. Based on the experimental results, the best model configuration is padding to a certain length that is not much different from the data, the extraction layer output topology is the same expression aspect and sentiment expression, RNN BiReGU cells, number of hidden units 250, dropout rate 0.5, and cross share k 5. F1-measure on the aspect and sentiment term extraction task for token and entity levels is 0.9081 and 0.9, better than the baseline model. However, for sentiment classification tasks the resulting model is still underperforming in most categories compared to the baseline.