Novel Approach to New Domain Aspect Identification Using Deep Learning and Word Replacement

Abstract

Customers of businesses enter a vast amount of data online. The data are analyzed by businesses to measure customer sentiment, allowing businesses to gain insights into what customers feel about their products and services. Sentiment analysis is therefore of prime importance to businesses. Sentiment predictions in sentences can be improved by identifying aspects, also known as Aspect Term Extraction (ATE), and then measuring the polarity of the sentiments associated with the aspects. This detailed measurement of aspect sentiment polarity helps in the overall sentiment classification of sentences. In the area of deep learning using neural networks, embeddings are created across a large corpus of data. These embeddings provide similarity measures that can be used for predicting aspects in unseen domains. In the proposed approach, unseen words of a domain are replaced with words from the training domain. The replacement words are identified using similarity scores between unseen words of a domain and existing words of the training domain. The modified sentences afterword replacement are given to the various domain-specific models for unseen domain aspect prediction. Extensive experiments are carried out across ten datasets to validate our approach. In a majority of the unseen domains, and across the different algorithms used, our approach performs well and gives a higher overall prediction accuracy.