In recent years, significant advances have been made in novel modelling tasks centred on aspectual sentiment analysis. This task shows relatively good results by extracting aspect-category-opinion-sentiment (ACOS) quaternions. To further improve the performance of the model in ACOS quadruplet extraction, this study proposes an integrated approach combining encoder-decoder networks, multi-level perceptrons (MLP), supervised contrast loss (SCL) and representation structure (ES) metrics. The integration of these techniques aims to improve the generation of more informative features by propagating aggregated representations from the encoder to the multilevel perceptron. In addition, contrast learning methods are used to encourage models to generate discriminable input representations on key attributes, thereby facilitating distinguishable input representations that facilitate the task of second-level prediction. In addition, an expression structure output target was introduced which was combined into an autoregressive encoder-decoder model to generate quartets, and empirical results confirmed the excellent performance of the model on three ACOS datasets. The experiment effectively established the feasibility of the task and highlighted its strengths in generating aspects for sentiment analysis.