Evaluation of end-to-end aspect-based sentiment analysis methods employing novel benchmark dataset for aspect, and opinion review analysis

Abstract

Aspect-based sentiment analysis (ABSA) deals with the determination of sentiments for opinion targets. While historically this research task has been addressed with pipeline approaches, more recent works use neural networks to jointly deal with the aspect term and opinion term extraction, as well as the polarity classification. Although learned together, most NN-based approaches and all pipeline approaches do not model correlations between the tasks. This is also based on the absence of adequate datasets which are annotated for all sub-tasks in a unified tagging scheme. We address this bottleneck and introduce the first purposely designed and annotated dataset for ABSA. The DAORA dataset covers 2,100 Tripadvisor reviews, and it is annotated on aspect terms, opinion terms, as well as aspect term polarity, using a unified tagging scheme. It was designed especially for end-to-end aspect-based sentiment analysis of real-world reviews and does not use any sentence repetition or removal. We evaluate the DAORA dataset in several experiments employing state-of-the-art models for ABSA. We set benchmarks and analyze the strengths as well as weaknesses of the data and approaches.