Aspect-Opinion Extraction with Polarity Estimation through Dependency Relation Analysis

Abstract

Opinion Mining (OM) or Sentiment Analysis (SA) emphasizes on the study of the customer’s behavior likewise; as attitude, requirements, and desires about a product or service. Aspect-based Sentiment Analysis (AbSA) provides an analysis of customers’ sentiments on different product/service aspects or features at a finer level, in the form of reviews posted on social media platforms.  In AbSA, the core task is to identify and extract product facets, ranking, and then classification. For this purpose, supervised, unsupervised, and semi-supervised methods are used for Aspect Extraction (AE). Moreover, several approaches and algorithms have been recommended in the literature. Dependency Relation Analysis (DRA), an AE method; uses Type Dependency Relations (TDRs) to extract significant product aspects linked with sentiments. For example, linguistic features serve as a mainstream backbone for language analysis. This study intends to extract aspect-opinion pairs along with their sentiments, applying an unsupervised approach by using DRs’ and rule-based algorithms. To evaluate the proposed system’s effectiveness, the APR dataset was used and results were compared with the baseline studies. The outcome from the proposed method demonstrates that it outperforms the baseline studies with 0.85, 0.75, and 0.79% for Precision, Recall, and F1-measure performance metrics, respectively. Besides it, the customer reviews polarity estimation was investigated at a large scale with an enhanced rule-based algorithm that results in improved effectiveness.