Enhancing Aspect-Based Sentiment Analysis: Investigating Aspect Term Extraction and Annotation Schemes

Abstract

We live in an age where there is constantly an increasing need for detailed information. From the company that needs to know what their users are feeling and thinking about their product to another crisis like a global pandemic, there is a growing need for information. We have seen various crises including mental health, natural disasters, and economic change. On the other hand, we desire the advancement of technology so that we can know more, experience more, and enjoy more. To achieve an improved experience, organizations need a way to understand their users. Computers aid in bridging the gap between organizations and their users’ needs. For computers to gather detailed sentiment information from text, there is a necessity for an approach that analyzes the content of the text. Typically, Sentiment Analysis (SA) is implemented at a document level, sentence level, aspect level, fine-grained level, and emotion level. There is a range of depth to the levels of SA. However, one of the most fine-grained levels of SA is Aspect Based Sentiment Analysis (ABSA). It is one of the standard most in-depth approaches to SA. While the ABSA approaches lead to detailed information, they continue to produce limits in information by solely focusing on the polarity of an aspect term. Thus, the outputs of ABSA models are deficient. We thus need to design and develop approaches for to improve ABSA which is why we focus on Aspect Term Extraction This dissertation argues that to provide detailed information from text, we must design techniques for ATE. In this regard, we design an ATE framework to address it’s integral part of the development of the sentiment analysis branch of Natural Language Processing (NLP). We believe that a successful ATE approach using annotation schemes has not yet been elucidated. In the first part of this dissertation, we provide the background of ABSA, why there is a need for ATE. We then provide one case study that exposes the limits of topic modeling as an approach to ATE in ABSA which has been previously used. Through our experiments, we develop a framework of Sentiment and Emotion Analysis (SEA) that presents the weakness of topic modeling that is commonly used for Aspect-Based Sentiment and Emotion Analysis. In the second part of this dissertation, we present a unique neural network-based approach to ATE using annotation schemes. Additionally, we present our own annotation scheme, which we all reverse IOB (or IOB), as a successful approach to Aspect Term Extraction. The method models aspect extraction from the data and the emotion analysis of the aspect text. The output of our models produces results that make our approach a viable solution for producing detailed information. We conclude our dissertation with an opportunity for ABSA to be a tool for detection in online tasks. This dissertation exhibits how to develop approaches that contain these properties and which implementations will return accurate information.