Enhancing Aspect-Based Sentiment Analysis with Contextualized Window Attention Mechanism

Hoang-Ha Do, Xuan-Hieu Pham, Duc-Hiep Nguyen, Quoc-An Nguyen, Duy-Cat Can, Hoang-Quynh Le

Abstract: In recent years, sentiment analysis has been a heated topic across the globe in general and particularly in Vietnam. Sentiment analysis plays an important role in promoting the economic growth. For instance, marketing departments from big corporations rely mostly on analytic of customers's attitude to develop marketing strategy for their company. Knowing its significance, we implement a system specifically for this problem. Most traditional approaches for Vietnamese sentiment analysis learn the context weight of every word in a sentence. However, we discover that in most cases only a part of a sentence, which consists of a group of continuous words actually influences the polarity of an aspect. Hence, we introduce a novel approach called Contextualized Window Attention, which helps our models learn the context of such groups instead of a single word. We experiment with 4 different models and train them on dataset crawled from Shopee and Tiki, which are 2 of the most popular and trusted E-commerce platforms in Vietnam. However, across every aspect, comments that are negative makes up of only 5% approximately of total comments. Hence, we only considerate it valuable when evaluating on the negative label. Our best model obtained the maximum micro- and macro-average performance of 76.66% and 78.62% on the negative label, respectively.