Fine-grained cybersecurity entity typing based on multimodal representation learning

Fine-grained entity typing is crucial to improving the efficiency of research in the field of cybersecurity. However, modality limitations and type-labeling hierarchy complexity limit the construction of fine-grained entity typing datasets and the performance of related models. Therefore, in this paper, we constructed a fine-grained entity typing dataset based on multimodal information from the cybersecurity literatures and design a multimodal representation learning model based on it. Specifically, we design and introduce a new benchmark dataset called CySets to facilitate the study of new tasks and train a novel multimodal representation learning model called Cyst-MMET with multitask objectives. The model utilizes multimodal knowledge from literature and external to unify visual and textual representations by eliminating visual noise through a multi-level fusion encoder, thereby alleviating data bottlenecks and long-tail problems in the fine-grained entity typing task. Experimental results show that CySets have sharper hierarchies and more diverse labels than the existing datasets. Across all datasets, our model achieves state-of-the-art or dominant performance (3%), demonstrating that the model is effective in predicting entity types at different granularities.