TUPLE OPINION EXTRACTION USING GENERATIVE APPROACH BASED ON PRE-TRAINED MODEL FOR ASPECT BASED SENTIMENT ANALYSIS

• Main Author: Athallariq Harya P, Danendra
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78026
• Institution: Institut Teknologi Bandung

The widespread use of social media as a means of conveying public opinion has made sentiment analysis an increasingly promising field of study in the future. Aspect-based sentiment analysis (ASBA) can be used by industry to view criticisms and suggestions from the public regarding aspects of the industrial services and products they offer. ASBA can be subdivided into smaller sub-tasks based on the identification it does. The large number of ASBA sub-tasks makes comparisons between previous studies difficult and also requires model adjustments due to specific designs to complete only certain sub-tasks. In this study a sub-task called opinion tuple extraction was designed which was a combination of the previous ASBA sub-tasks. From an input text the model will predict opinion tuple pairs consisting of (entity, aspect, opinion, sentiment, aspect category) for each aspect in it. Opinion tuple extraction can be accomplished with a generative approach based on pre-trained language models and other ASBA sub- tasks can be accomplished by a simple decoding process of opinion tuple extraction labels. Opinion tuple extraction has the advantage of flexibility in solving various ASBA sub-tasks because it only requires the decoding process of the labels compared to previous research which required adjustments to the model architecture. In this thesis research, we tested the different effect of text pre- processing, pre-trained language models used, post-training, and post-processing with Levenshtein distance on the performance of opinion tuple extraction. The best model has precision performance of 0.836, recall of 0.833, and f1-score of 0.834 which was obtained using standard text cleaning (removing urls, mentions, emojis and hashtags) plus the use of stemming, using the IndoT5 pre- trained language model which was post-trained to improve domain knowledge and task awareness and with post-processing. This model outperforms the performance of the baseline model from previous studies.