Target-Aspect-Sentiment Joint Detection: Uncovering Explicit and Implicit Targets Through Aspect-Target-Context-Aware Detection

Abstract

Target Aspect Sentiment Detection (TASD) is challenging because it involves various Natural Language Processing (NLP) subtasks including opinion target detection and sentiment polarity classification. Despite significant advancements in this area, most studies have neglected the interrelation between opinion elements and contexts, primarily when a target opinion is expressed implicitly. This study proposes Aspect-Target-Context-Aware Detection for Target Aspect Sentiment Detection, which is a joint learning neural-based framework. The Aspect-Target-Context-Aware Detection model incorporates opinion context syntactic information by utilizing dependency relations associated with opinion terms, and considers head nodes as a primary element for identifying relevant opinion contexts. The Target Aspect Sentiment Detection task was divided into aspect sentiment classification and opinion target extraction tasks. For aspect sentiment, multiclass classification was employed for aspect-sentiment pairs. A BIO tag inference scheme is adopted to detect the opinion target and determine its type (implicit or explicit) for opinion target extraction. The approach was evaluated using two restaurant datasets: Task-5 of SemEval-2016 and Task-12 of SemEval-2015. The proposed approach demonstrated cutting-edge performance when extracting multi-opinion elements from the TASD task, with notable improvements in Macro-F1 values: 3.28% for SemEval 2015 and 5.97% for SemEval 2016. The model also identifies various opinion types and offers valuable insights for future developments, particularly for implicit opinion detection.

Keywords

• Aspect-based sentiment analysis (ABSA)
• dependency relations
• explicit opinion
• implicit opinion
• target-aspect-sentiment (TASD)