An aspect-opinion joint extraction model for target-oriented opinion words extraction on global space: An aspect-opinion joint extraction model

In aspect-based sentiment analysis, target-oriented opinion words extraction (TOWE) aims to extract opinion words based on aspect terms. Most current methods used in TOWE tasks only focus on explicit aspects and tend to overlook the implicit aspects, leading to a bias in the sample selection process and incomplete modeling of the TOWE tasks. Therefore, it is essential to consider both explicit and implicit aspects simultaneously in the modeling process of TOWE tasks. This paper proposes an aspect-opinion joint extraction (AOJE) model composed of an aspect term extraction unit (ATEU) and a target-oriented opinion words extraction unit (TOWEU). ATEU first is responsible for extracting aspect terms and converting them into prompt templates. TOWEU uses these templates to obtain opinion words for specific targets. This model is trained and evaluated on global space, including explicit and implicit aspects. This approach effectively addresses the issue of sample selection bias. The proposed AOJE method performs better than existing methods by an average of 4.06% on the Macro-F1 score on the SemEval 14-16 datasets. In particular, the AOJE model shows significant improvements compared to the IOG (Inward-Outward LSTM+Global context) model, with Macro-F1 scores increasing by 9.00%, 8.48%, 7.41%, and 9.21% on the Laptop 14, Restaurant 14, Restaurant 15, and Restaurant 16 datasets, respectively. These experimental results indicate that the AOJE model trained on global space significantly enhances the performance of TOWE and improves generalization capabilities.