<p>To resolve the limitations of using a singular representation of the adjacency matrix in contemporary sentiment analysis with graph networks, as well as the inadequate incorporation of syntactic structure and lexical semantics in triplet extraction for sentiment analysis, we propose SenticNet-based Triplet Aspect-level Sentiment Analysis (Sen-TriSA) model. The process begins with vectorizing words and their part-of-speech (POS) information using the Bidirectional Encoder Representations from Transformers (BERT) embedding model. In order to extract lexical semantic information, the combined word and POS vectors are then fed into a Bidirectional Long Short-Term Memory (BiLSTM) layer. The word vectors and the enhanced adjacency matrix, enriched with sentiment scores from the SenticNet lexicon, are simultaneously input into the Graph Convolutional Network (GCN) to extract syntactic structural information. Then, vectors derived from a self-attention mechanism and lexical semantic information are fused with the syntactic structural information. Subsequently, the aspect categories and sentiments are identified using a softmax function, while aspect terms are extracted with a Conditional Random Field (CRF) model. This process achieves the extraction of aspect category–aspect term–sentiment triplets, enabling finer-grained aspect-based sentiment analysis. The comparative and ablation tests conducted on publicly accessible datasets demonstrate the effectiveness and robustness of the proposed model.</p>