LongMuEL: An entity disambiguation method targeting multiple sources of data and long text

Entity disambiguation is an important method for completing knowledge graphs. The data used for knowledge graph completion often integrates various data types and formats. When designing models for entity disambiguation, a key task is to design a knowledge fusion module to efficiently obtain embeddings of candidate entities from the descriptive text and structured data of the candidates. Since the length of candidate entity descriptive text usually far exceeds the limit of input text for pre-trained models, how to incorporate truncated text knowledge into the embeddings of candidate entities is also a problem to be considered when designing entity disambiguation models. In this paper, we propose a novel entity disambiguation model, LongMuEL, which can represent the long descriptive text of candidate entities and integrate structured text from candidate entity knowledge bases. In the long text embedding module, we use multiple attention modules and weight pooling for all truncated text. When obtaining the embedding of structured text, we adopt two self-supervised tasks to pre-train the pre-trained model and generate negative examples through negative sampling. We demonstrate that pre-training the model with two self-supervised tasks results in better embedding effects for structured data. Finally, we fuse the embedding vectors generated from the two data types to calculate the ranking of candidate entities. LongMuEL is a knowledge graph completion model adapted to existing data features, and through ablation experiments, we verify the effectiveness of each part of LongMuEL.