Efficient Ultrafine Typing of Named Entities

Ultrafine named entity typing (UFET) refers to the assignment of predefined labels to entity mentions in a given context. In contrast to traditional named entity typing, the number of potential labels is in the thousands and one mention can have more than one assigned type. Previous approaches either depend on large training datasets, or require inefficient encoding of all input-type combinations. Therefore, there is a need for investigating the efficiency during training and prediction of entity typing models in the ultrafine-grained setting, considering its distinctively bigger search space, compared to the coarse- and fine-grained tasks. To efficiently solve UFET, we propose Decent, a lightweight model that encodes, using a pretrained language model, the input sentences separately from the type labels. Additionally, we make use of negative oversampling to speed up the training while improving the generalization of unseen types. Using an openly available UFET dataset, we evaluated the classification and runtime performance of Decent and observed that training and prediction runtime is orders of magnitude faster than the current state-of-the-art approaches, while maintaining a competitive classification performance.