From Implicit to Explicit: A Simple Generative Method for Aspect-Category-Opinion-Sentiment Quadruple Extraction

Aspect-Category-Opinion-Sentiment Quadruple Extraction (ACOS) is a critical subtask of Aspect-Based Sentiment Analysis (ABSA), aiming to extract all quadruples <aspect, category, opinion, sentiment> in a review sentence. Existing ACOS methods are categorized into pipeline methods and unified methods. Pipeline methods face error propagation and ignore the interdependency among four sentiment elements. Unified methods generate a long target sequence when the number of quadruples increases, which degrades the model performance. In this paper, we pay attention to the implicit aspects and opinions to obtain comprehensive aspect-level sentiment information. To this end, we propose a novel sequence generation model for ACOS named SG-ACOS. We design a linearization method to express ACOS quadruples as a sequence. The proposed linearization method incorporates natural language tokens and two special tokens into the target sequence. Natural language tokens are employed to represent the four sentiment elements, thus facilitating the model to learn the semantics of sentiment elements. Special tokens reduce the length of the target sequence, thereby improving the model efficiency and performance. Our proposed model obtains an absolute F1 improvement of 3.52% and 2.74% against previous state-of-the-art methods on two ACOS datasets, respectively. Further experimental results show the model effectiveness in implicit sentiment detection and the robustness of our model.