<p>Multimodal sentiment analysis (MSA) is an important task that predicts human sentiment by fusing data from different modalities, including text, audio, and images. Although modern deep learning approaches, including graph convolutional networks and transformers, have achieved promising results, several challenges remain, particularly in effectively modeling complex inter-modal relationships, learning comprehensive feature representations, and addressing modality imbalance during information fusion. To address these problems, this paper proposes an architecture that combines graph convolutional networks (GCNs), transformers, and contrastive learning (called GCN-TCNet) to improve the performance of sentiment analysis from text + audio by leveraging the contextual relationships between modalities. The proposed model includes the following main steps: (i) representing and extracting features from text and audio using two encoders, RoBERTa and Wav2Vec2; (ii) enriching the extracted feature representations with important contexts from texts and audio by constructing two GCNs based on the combination of GAT and SAGE layers; (iii) fusing two enriched representations using the transformer architecture with multi-head attention mechanism; (iv) determining the sentiment polarity of opinions expressed in text combined with audio by constructing a classifier based on the multiple layer perceptron technique; (v) training the GCN-TCNet model by optimizing the combined loss function of MSE, smooth L1 and focal loss and applying contrastive learning technique to enhance the compatibility and complementarity between the modalities. The GCN-TCNet model is tested on benchmark datasets, and the results indicate that the proposed model achieves state-of-the-art performance on the CMU-MOSI (Corr = 0.816) and CH-SIMS (Corr = 0.640) datasets, and remains highly competitive on the CMU-MOSEI dataset.</p>