Aspect Based Sentiment Analysis by ensembling Pre-trained BERT models

Abstract

Sentiment Analysis has been becoming an important task for academics as well as for commercial companies. However, most of the current approaches target iden tifying the overall polarity of a sentence instead of identifying the polarity of each aspect mentioned in the sentence. Aspect Based Sentiment Analysis (ABSA) deals with identifying the aspects within the given sentence and the sentiment expressed for each aspect. ABSA expects a single entity per given sentence. In Targeted Aspect Based Sentiment Analysis (TABSA), the goal is to identify the sentiment polarity towards each aspect of one or more entities. Recently, the use of pre-trained models such as BERT, has resulted in achieving state-of-the-art results in the field of natural language processing. In this work, we propose two ensemble models based on BERT, namely BERT-E-MV and BERT-E-AS. We construct an auxiliary sen tence from the aspect using different methods and convert (T)ABSA problem to a sentence-pair classification task. We then fine-tune different pre-trained BERT mod els and ensemble them for final prediction and achieve new state-of-the-art results on datasets belonging to different domains in English as well as Hindi languages.