Aspect Based Sentiment Analysis Berbasis BERT dan Multi-Task Learning
Muhammad Naufal Hakim, Syukron Abu Ishaq Alfarozi, S.T., Ph.D.; Prof. Ir. Paulus Insap Santosa, M.Sc., Ph.D., IPU.
2024 | Skripsi | TEKNOLOGI INFORMASI
The rapid development of machine learning has facilitated the categorization of text into positive, negative, and neutral classes through sentiment analysis models. However, sentiment analysis still lacks detail in capturing specific aspects of a topic. To address this, prior research has developed Aspect-Based Sentiment Analysis (ABSA) to discern sentiment based on aspects or topics within a text. The most recent advancements in ABSA have leveraged multi-label learning principles, utilizing a classifier for each aspect labels present in the dataset, where the classification of both aspect and sentiment on a given label is conducted concurrently within a single classifier.
This research endeavors to explore the application of multi-task learning in Aspect-Based Sentiment Analysis (ABSA) models, utilizing two classifiers for each label, one specifically for aspect classification and the other for sentiment classification. Consequently, each classifier in the model will possess enhanced specialization, being focused on a more specific task. Moreover, the model will learn from the combined loss values of aspect and sentiment, this presents a potential for performance enhancement by adjusting weight values between aspect and sentiment classification tasks.
The findings of this study indicate that, the multi-task learning model (F1 = 96.16%) surpasses the non-MTL model (F1 = 94.78%) in the combined classification of aspects and sentiments. In separate metrics between aspects and sentiments, MTL excels in aspect classification (F1 = 96.93% vs 96.33%) and sentiment classification (F1 = 94.13% vs 90.33%). The loss weighting demonstrates that optimal performance for the aspect task and sentiment task is achieved when a heavier weight is placed on aspect classification (alpha = 0.55), underscoring the importance of aspect accuracy is needed for precise sentiment prediction.
Kata Kunci : aspect based sentiment analysis, multi-task learning, multi-label learning, BERT, loss weighting