Neural Zero-Shot Fine-Grained Entity Typing

Abstract

Fine-grained entity typing is a task to assign types to entity mentions dependent on mentions’ context. Due to the heavy work of human annotation, high quality training data is always not enough, zero-shot fine-grained entity typing becomes important. Previous zero-shot works rely on hand-crafted features and suffers from noisy distant supervision induced training data. In this paper, we propose a Neural Zero-Shot Fine-Grained Entity Typing (NZFET) model. NZFET is an end-to-end neural model free from hand-crafted features. The entity type attention in NZFET makes model focus on information relevent to the entity type. In our experiments, NZFET obtains better results on popular datasets than previous works. And we show experimentally that our method is robust against noisy training data.