SI-MEL: A dataset of multi-modal entity linking based on scientific information

<p indent="0mm">Entity linking is one of the core tasks in natural language processing, and plays a crucial role in a wide range of downstream applications, such as information extraction, knowledge graph construction, and question answering. In the field of scientific information, however, due to the lack of large-scale, high-quality public datasets have led to notable performance limitations. Especially when dealing with phrases such as technical terms and abbreviations, these models often fail to effectively capture semantic relationships, resulting in suboptimal linking accuracy. To address this issue, this study designs a data processing framework based on Wikipedia articles related to scientific information. The framework includes raw data acquisition (article crawling and data cleaning), candidate entity generation (mention extraction and retrieval, image crawling, similarity computation), and final dataset construction (manual verification). Using this framework, we constructed the Science Information - Multimodal Entity Linking ( SI-MEL), a multimodal entity linking dataset covering three main themes: organizations (1,415 entries), people (4,932 entries), and specialties (8,693 entries). Experimental results demonstrate that, through various input expansion strategies, the enhanced structure and diversity of the SI-MEL dataset significantly improve entity linking performance. The release of SI-MEL provides an important research resource for multimodal entity linking in the field of scientific information and holds substantial significance for advancing research in this area.