Recent Advances and Future Directions

Since around 2015, many approaches for entity linking started to leverage neural network or deep learning models, pushing the state-of-the-art performance on this task. We have introduced some of them in the previous chapters. In this section, we highlight two recent threads of work: dense entity retrieval and auto-regressive entity retrieval. Entity linking can be viewed as an information retrieval problem. The query is a mention in text, and the goal is to retrieve a Wikipedia article that the mention refers to. Traditional retrieval methods such as TF-IDF and BM25 are built on sparse representations of both queries and documents. Recently, several works have shown that using dense representations could achieve a superior performance on a range of retrieval tasks (Reimers and Gurevych 2019; Gillick et al. 2019; Hofstätter et al. 2021).