An Entity Disambiguation Method for Chinese Short Text Based on Multi-Task Learning

As a fundamental task in natural language processing, entity disambiguation aims to remove the mention's ambiguity by linking it to target entity in the knowledge base. Existing methods have low accuracy due to the problems of multiple meanings of words and the semantic sparsity of entities in Chinese short texts. In this paper, we propose an entity disambiguation method based on multi-task learning for Chinese short texts with limited contextual information and informal expressions, which assists entity disambiguation with entity type so that the model can learn deeper semantic expression information. Firstly, concatenate short texts with data from the knowledge base to form an extended text with enhanced contextual information. On this basis, to make full use of key features in the short text, we add an adaptive maximum pooling layer to capture the semantic information of short texts, thereby optimizing the performance in entity disambiguation. Furthermore, for the issue of loss weight allocation in multi-task learning, we propose three different strategies. In this paper, experiments are conducted on two datasets, CCKS2019 and CCKS2020, and the F1 score metrics are 0.9455 and 0.9393, respectively. This indicates that auxiliary tasks and adaptive maximum pooling can alleviate the problem of insufficient contextual information in short texts and improve the model performance; the weight allocation strategy can lead the model to give different attention to sub-tasks during training, affecting the final performance of the model.