[Defense] Joint Detection of Aspect-Based Sentiment Analysis using Extended Review Context
Thursday, December 15, 2022
1:00 pm - 2:45 pm
Siva Uday Sampreeth Chebolu
will defend his proposal
Joint Detection of Aspect-Based Sentiment Analysis using Extended Review Context
User-generated reviews have become an integral part of our everyday life. Due to the exponential increase in the number of reviews generated everyday, it is now more complex to read all reviews for every product, location, or business option. In such a situation, it is crucial to automatically summarize the opinions from the reviews in a form that is meaningful and easily understood by everyone. In addition, internet review trends have shifted from assessing a product’s overall performance or quality to evaluating its major aspects. The analyses and methods used to summarize reviews should also be updated to keep up with user needs and current trends. Aspect-based sentiment analysis (ABSA) is a Natural Language Processing (NLP) task that involves analyzing user-generated reviews to determine (i) the target being evaluated, (ii) the aspect category to which it belongs, (iii) the opinion expression, and (iv) the sentiment expressed towards any combination of the elements above. Current ABSA systems have proposed several subtasks for diverse applications to approach ABSA from a broad range of perspectives. These subtasks have identified only a subset of ABSA elements, leaving a few essential relationships to collectively identify them to develop a holistic system capable of delivering every conceivable summary information. In addition, most proposed techniques tackle the ABSA task at the sentence level as opposed to analyzing the complete review, hence missing critical context information accessible from the prior or following sentences. In this proposal, we address the shortcomings mentioned above by providing novel and efficient approaches for jointly detecting all ABSA elements at both the sentence and review levels. First, we examine the elements’ inter-dependencies to discover the inherent relationships that may aid in a joint formulation. Then, as a step toward the primary objective, we present an intuitive method for combining these segments using a novel task formulation. Furthermore, we plan to examine how the sentence-level ABSA systems generalize across the entirety of the reviews, given the additional context offered by other sentences in the review. In addition to the proposed methods, we aim to provide benchmark datasets for joint detection at both the sentence and review levels to test and analyze them in the context of the primary research objective.
1:00PM - 2:45PM CT
Online via Zoom
Dr. Thamar Solorio, dissertation advisor
Faculty, students, and the general public are invited.