Chinese Short Text Entity Linking Based On Semantic Similarity and Entity Correlation

Nowadays, more and more corpora in Internet consist of short texts, such as QA queries, posts in social media and news titles. Entity linking for these short texts is quite important. However, due to the serious colloquialism and insufficient contexts, the entity linking task of Chinese short text is more difficult. In this paper we propose an entity linking model named BERT-RWR. BERT-RWR aims at improving the accuracy of predicting the right target entity in candidate set by integrating deep neural network and graph model. More specifically, deep neural network based on fine-tune BERT is designed to calculate the mention-entity semantic similarity and Random Walk with Restart (RWR) algorithm can further capture the correlation between candidate entities of different mentions. In BERT-RWR, we leverage (1) semantic similarity score between each mention and its candidate entities and (2) the prior probability and (3) the correlation between different candidate entities to select the target entity. To improve the recall rate of candidate entity, we put forward three-method fusion strategy for candidate generation. Experimental results demonstrate that our model outperforms the state-of-the-art results for entity linking in Chinese short text datasets.