A novel end-to-end neural network for simultaneous filtering of task-unrelated named entities and fine-grained typing of task-related named entities

Recently, one emerging problem in Named Entity Typing (NET) is the fine-grained classification of task-related entities co-existing with task-unrelated entities. The traditional pipeline framework decomposes this problem into two sub-tasks. The first sub-task filters out the task-unrelated entities, while the second sub-task performs fine-grained classification for task-related entities. In the present study, we have developed an end-to-end neural network to solve the two sub-tasks simultaneously. The new model has two main merits. First, Mention–Mention (MM) relationship learning is developed to capture the interaction of task related and unrelated entities for producing more discriminative features. Second, an Improved Radial Basis Function classifier (ImRBF) with a novel training scheme is developed to jointly solve task-unrelated entity filtering and fine-grained classification of task-related entities. Experiments show that our model outperforms the pipeline methods by 3.3%–6% (F1 score) on the first sub-task and 1.8%–6.3% (F1 score) on the second sub-task. •Task related and unrelated entities have very different properties.•Simultaneously recognizing all entities is challenging.•A novel end-to-end neural network is developed to classify all entities.•An improved RBF classifier applies closed decision boundary for all the entity typing.•Mention–mention relation strengthens the link of task related and unrelated entities.