Graph convolutional network with reinforced dependency graph and denoising mechanism for sarcasm detection

The widespread presence of sarcasm in social media presents significant challenges to sentiment analysis and public opinion monitoring, making accurate sarcasm detection particularly important. Recently, studies have found that graph-based methods can effectively overcome the limitations of sarcasm methods in modeling syntactic and emotional information. However, existing approaches often consider dependency relationships between nodes as homogeneous, and directly aggregating all edge information during the graph learning process can introduce noise. To address these issues, we propose a novel framework based on the graph convolutional network for sarcasm detection. This framework effectively improves the ability of sarcasm detection by modeling the heterogeneity of dependencies and introducing a noise suppression mechanism. Specifically, when constructing dependency graphs, we account for the importance of node distances and dependency types. To reduce the impact of noise, we extract robust feature representations using a denoising autoencoder. Then, a parameterized mechanism is applied to penalize sparse edges, suppressing the interference of irrelevant edge information. To evaluate the performance of the proposed model, we conducted experiments on three datasets of different scales, achieving F1-scores of 84.97%, 71.11%, and 83.83% respectively. The proposed model outperforms other baseline models across all metrics, demonstrating outstanding performance.