An Aspect based sentiment analysis model applicable for a domain with a small labeled data set

Abstract

Aspect Based Sentiment Analysis (ABSA) which is a more fine-grained approach to sentiment analysis has become a popular trend in the sentiment analysis field in academic as well as industrial contexts. It is beneficial for businesses and organizations as they can have a more detailed analysis of customer’s sentiments on each aspect of their service/product. However, training an ABSA model on a domain would be difficult when there is not enough labeled data. The goal of this research was to develop a deep learning model that can be utilized for ABSA on any domain with a small, labeled data set. The task was done under two sub-tasks, aspect extraction and aspect sentiment extraction. Several approaches to performing these two sub-tasks under limited labeled data were compared with a Convolution Neural Network (CNN) as the base model. From the approaches tested, the multichannel CNN model with the data augmentation approach outperformed all the other models with an accuracy of 73% for aspect extraction and 78% for aspect sentiment extraction. This proposed model was able to achieve good accuracy even with very limited labeled data, compared to the benchmark that used a dataset three times larger.