imEL: Instance-level Masked Entity Linking Model

Knowledge base question answering (KBQA) is a complicated natural language processing task. KBQA systems are used to answer questions about a large data set where the goal is to identify the entities in the question and link them to other relevant information in the same knowledge base. Prior entity linking (EL) systems inherently use a class-based approach whereby two similar entities distinguishable only through context are treated as a single entity. For example, two vehicles both mentioned as "blue SUVs" but with different time and location contexts will be treated as the same entity. This work introduces a new type of EL called instance-level EL that can recognize and utilize instance-specific information. One such instance-level EL model, the instance-level masked EL model (imEL) is described and evaluated. imEL has a high level of accuracy in responding to questions that require identification and linking of specific instances in the input text.