Transfer Learning for Aspect Term Polarity Determination

Abstract

Social Media and E-commerce have led to extensive growth in the amount of data collected as
feedback from people, related to product, entity or person of interest. This feedback, in the form of online
reviews, is analyzed at different levels. The sentence based and document based sentiment analysis are coarsegrained. However, the sentiment analysis done at a deeper level is Aspect Based Sentiment Analysis (ABSA).
Among the four subtasks of ABSA, we focus on the second, Aspect Term Polarity subtask. Aspect Term
Polarity subtask is concerned with determining the sentiment associated separately for each aspect term in the
review sentence in terms of positive, negative, neutral or conflict. In this paper, we target this subtask for HindiABSA dataset. Hindi, being a resource-scarce language, the Transfer Learning based method which involves
fine-tuning ULMFiT and MultiFiT models, is being proposed. The system shows an improvement in percentage
accuracy of 2.89%, over that of best reported state-of-the-art models. T