New GRU-CAP Method Integrating Sarcasm Detection Enhances Sentiment Analysis of Product Reviews

The Smart Mobility and Low-Altitude Economy Research Team has made important progress in the field of product feature sentiment analysis based on online reviews. The team has proposed an innovative method that integrates Gated Recurrent Units (GRU) and Capsule Neural Networks (CAP), incorporating the function of Chinese sarcasm recognition.

The study was published in the internationally renowned journal Expert Systems with Applications, which focuses on the research fields of artificial intelligence and decision science.

Against the backdrop of the rapid development of e-commerce and social media, product online reviews have become a core data source for enterprises to grasp consumer needs and optimize product design due to their large volume and strong authenticity. However, existing research on product feature sentiment analysis still faces significant limitations: on the one hand, it struggles to comprehensively extract explicit noun-based features and implicit features such as adjectives or phrasal verbs, with traditional methods exhibiting poor generalization and high computational complexity; on the other hand, the prevalent use of sarcasm in online language has not been adequately addressed, leading to high misjudgment rates in sentiment polarity. The unique advantages of GRU in processing sequential data and CAP in mining feature relationships and semantic recognition offer technical potential to address these challenges. Accordingly, this study innovatively proposes a GRU-CAP dual-channel parallel architecture model. The model first employs the GRU-CAP feature extraction module to extract both explicit and implicit product features from reviews and generate corresponding feature indices. It then integrates the GRU-CAP sentiment analysis module to simultaneously perform sentiment polarity classification and sarcasm semantic recognition, ultimately achieving fine-grained product feature sentiment analysis. Using mobile phone reviews from the JD.com platform as a case study, experiments validate the accuracy and effectiveness of this method in both feature extraction and sentiment classification. The overall F1-score reaches 96.38%, significantly outperforming traditional machine learning and single deep learning models.

Paper link:

https://doi.org/10.1016/j.eswa.2023.122512