Abstract: To address the computational redundancy caused by multi-stage partitioning in pipeline strategies and the high resource consumption and low computational efficiency resulting from complex tagging designs in end-to-end joint extraction strategies for aspect sentiment triplet extraction, this paper proposes a lightweight method based on triplet loss, termed BTLL-ASTE (Based Triplet Loss Lightweight Grid Tagging Scheme Aspect Sentiment Triplet Extraction). In this method, a text encoder is first employed to obtain contextual semantic features of sentences, and a BiGRU module is further introduced to enhance the discriminative power of these features. The enhanced features are then fed into both the triplet joint loss (TJL) component and the grid tagging strategy (GTS) component. Within the TJL and GTS components, layer normalization is applied to facilitate deep interaction between aspect and opinion features. Meanwhile, a weight balancing factor is incorporated to increase the penalty on sparse categories, thereby effectively alleviating the class imbalance problem and optimizing model training to improve the quality of inputs to the decoder. Finally, the decoder combines the optimized model with predefined semantic rules to output complete aspect sentiment triplets. Comparative experimental results on four public datasets demonstrate that the proposed method achieves an average improvement of 2.89% in precision, verifying its effectiveness in reducing resource consumption and enhancing extraction accuracy.