OPINION TRIPLET EXTRACTION FOR ASPECT BASED SENTIMENT ANALYSIS USING GRID TAGGING SCHEME

• Main Author: Pradipta Wirawan, Gama
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/56237
• Institution: Institut Teknologi Bandung

Aspect-based sentiment analysis (ASBA) is one of the variations of sentiment analysis that can be used by companies to find out public opinion in detail on aspects related to the products or services provided. There are several subtasks under ASBA, namely aspect/sentiment term extraction, aspect categorization, extraction of aspect and sentiment terms relations, and sentiment classification. Opinion triplet extraction is a combination of several previous subtasks, which aims to extract three opinion factors from the review sentence (aspect expression, sentiment expression, sentiment polarity). In general, these tasks are performed separately sequentially. However, this approach is considered less efficient and has the potential to reduce model performance due to errors in the previous process. The Grid Tagging Scheme approach performs opinion triplet extraction simultaneously and gets better performance than the pipelined approach. In addition, this approach can also overcome one of the problems in extracting aspect sentiment pairs, namely overlapped triplet, which means that there are one or more aspect terms that have two or more different opinion terms, and vice versa. This final project focuses on adapting this approach for extracting triplet opinion from Indonesian hotel reviews. Based on the experimental results using the Airy dataset, the best model configuration is to include incomplete triplet data into the training data, use a monolingual language model and use a fine-tuning strategy in the model training process. The F1-score of the opinion triplet extraction task is 0.78. As for the aspect expression and sentiment expression extraction tasks, the F1-score of the test is 0.87, which is lower in performance than the baseline model.