Knowledge-Based False Information Detection Model

In response to the problem of insufficient attention given by traditional detection methods to text structure, this article proposes four text relation graphs: sentence-entity graphs, syntax dependency-based graphs, semantic dependency-based graphs, and graphs that combine syntax and semantic dependency. Furthermore, the constructed text relation graphs are analyzed for their features, and it is found that the graph density of real texts is relatively high, indicating that the text structure in real texts is denser, and the number of entities and relationships is also richer. The analysis suggests a close correlation between text structure and false information, and considering text structure relations can help improve the accuracy of false information detection. To effectively aggregate the semantic features of text with multiple relations, this article proposes a graph neural network false information detection model based on multiple relation semantic aggregation. From the perspective of spatial topology and sequence relationships, GCN-LSTM and GAT-LSTM and other models with multiple relation semantic aggregation are constructed. Experiments compare different networks and topologies, different models, model fusion analysis, and the results of each experiment are compared. Finally, the experimental results show that the GAT-LSTM training model with a syntax-semantic relation graph as the text feature input performs best, with an accuracy rate of 80.060%, recall rate of 81.843%, and F1 value of 82.520%.