Owner named entity recognition in website based on multidimensional text guidance and space alignment co-attention

In recent research, the task of Owner Named Entity Recognition (ONER) in websites has been proposed as a specific and practical application of Multimodal Named Entity Recognition (MNER). The ONER aims to identify the true owner of websites on the Internet, which plays a crucial role in network security. The existing method involves identifying the website owner’s name through the text, image, and domain in the content of the website, where the owner information usually appears. However, most of the previous methods simply extracted features from the image and the domain as two independent modalities and did not fully utilize the text information in them. Additionally, these methods do not consider that different modality features are trained on their respective modality space, which makes it difficult to model cross-modal interactions due to different feature spaces. To address these two issues, this paper proposes a Multidimensional Text Guidance and Space Alignment Co-Attention (MTGSAC) model to realize owner named entity recognition in websites. The MTGSAC model can utilize the text information in the image and the domain modalities to guide the text modality for features extraction. Meanwhile, the model designs a features fusion module based on Transformer and co-attention gate mechanism to effectively model cross-modal interactions. Furthermore, to address the problems of insufficient data samples and poor data diversity in the existing ONER dataset, we extended the ONER dataset and proposed the ONER-2.0 dataset. Experimental results on both the ONER and ONER-2.0 datasets show that our model achieves state-of-the-art performance.