Master of Science
Online text-based reviews are often associated with only an aggregate numeric rating that does not account for nuances in the sentiment towards specific aspects of the review's subject. This thesis explores the problem of determining review scores for specific aspects of a review's subject. Specifically, we examine two important subtasks - aspect identification (identifying specific words and phrases that refer to aspects of the review subject) and aspect-based sentiment analysis (determining the sentiment of each aspect). We examine two different models, conditional random fields and an association mining algorithm, for performing aspect identification. We also develop a method for performing aspect-based sentiment analysis based on VADER, a sentence-level sentiment analysis algorithm built for sentiment analysis of social media. We identify key problem considerations, including other important subtasks and ideal training dataset qualities, for future development of a full aspect-based review system.
Byrne, Sean, "Aspect Identification and Sentiment Analysis in Text-Based Reviews" (2017). Theses and Dissertations. 2535.