Fine-Grained Entity Typing via Label Noise Reduction and Data Augmentation

Abstract

Fine-grained entity typing aims to assign one or more types for entity mentions in the corpus. Recently, distant supervision has been utilized to generate training data. However, it has two drawbacks. First, the same labels are assigned to every entity mention in a context-agnostic manner, which introduces label noise. Some approaches alleviate this issue by hand-crafted features. However, they require efforts from experts. Second, the entity mentions out of Knowledge Base (KB) are ignored and hence cannot be added to the training data, which decreases the size of the training data. Furthermore, the existing entity typing systems neglect the types of other entity mentions in the same context which provide evidence to infer the types of the target entity mentions. In this paper, we first propose graph-based and sampling-based approaches, to reduce the label noise generated by the distant supervision, and then augment the training data by finding potential entity mentions in the corpus and inferring their types. Moreover, we propose a hierarchical neural network, which involves the types of other mentions in the context and satisfies the type consistency, to predict the types. Experiments on two datasets show that our system outperforms state-of-the-art entity typing systems.