A survey on aspect sentiment triplet extraction methods and challenges

Abstract

Aspect-based sentiment analysis (ABSA) has gradually become an important technique for mining online reviews and is widely popular across various domains, such as producer–consumer, pharmaceutical reviews, political campaigns, and celebrity popularity. Aspect sentiment triplet extraction (ASTE) is a core technique within the ABSA, as it automatically extracts aspect terms, opinion terms, and sentiment polarity triplets from textual data. Since the ASTE task is a relatively recent research direction, there is still a lack of comprehensive summaries and syntheses of the research in this task. To address this issue, this paper provides a comprehensive introduction to various methods, performance evaluations, challenges, and future research directions for the ASTE task. Specifically, we categorize the current ASTE approaches into six types: Pipeline, End-to-end, Generative, MRC-based, Table-filling, and Span-based methods. Subsequently, we provide a detailed introduction to the characteristics of each method, along with their strengths and weaknesses. Additionally, we organize the performance of these methods on two benchmark datasets, ASTE-Data-v1 and ASTE-Data-v2. Finally, we discuss the challenges faced in current work and potential future directions.