Semi-Supervised Learning for Aspect-Based Sentiment Analysis
Student thesis: Doctoral Thesis
Related Research Unit(s)
Aspect-Based Sentiment Analysis (ABSA) is to find the aspects and their corresponding sentiments of a reviewed entity (e.g., a product or service) in various opinion documents (e.g., user reviews), which enables more fine-grained analysis on the opinion posted by people. ABSA normally involves two main sub-tasks, namely Aspect Mining (AM) and Aspect Sentiment Classification (ASC). AM sub-task aims to extract the words describing aspects of a reviewed entity, and the ASC sub-task is to analyze the expressed sentiments on the aspects. With the rapid application of deep learning technologies, supervised models have recently achieved the best performance on both subtasks. However, training these deep models requires a large number of labeled texts which are very costly or unavailable, and they usually perform only one of the two sub-tasks, which limits their practical use. To this end, in this thesis, we present our research on exploiting semisupervised learning approaches to address AM and ASC sub-tasks, which can avoid the heavy requirement on the manual labeling on training data. Moreover, we also propose a new semi-supervised approach to finish the completed ABSA within a unified framework, thus the two sub-tasks can be addressed in an end-to-end manner.