OPEDABSA: A Dataset for Open Domain Aspect-Based Sentiment Analysis from Public Reviews

Anonymous

Abstract: Sentiment Analysis is core to customer management, product development and service delivery. In recent years, the need for Aspect-Based Sentiment Analysis (ABSA) has led to three shared tasks in SemEval (2014, 2015 and 2016), which attracted a large number of submissions from around the globe. Two challenges confronting ABSA are- low amount of data and constrained domain coverage. This work attempts to address these problems by presenting an open domain gold standard dataset (covering 111 fine-grained domains) curated from publicly available reviews. Along with the dataset, we also present strong baselines for four tasks- Aspect Term Extraction, Aspect Polarity Classification, Sentence Polarity Classification and End-to-End ABSA. We provide experimental results which show that our dataset helps models achieve a much better performance (~ 18.33% absolute improvement, on average) in open domain ABSA tasks.

Paper Type: long