Automatic Entity Typing using Deep Learning

Abstract

Knowledge bases contain vast amounts of information about entities and their semantic types. These can be leveraged in a variety of information access tasks like natural language processing and information retrieval. However, knowledge bases are incomplete, emerging entities need to be typed correctly, and existing entities must keep up to date. This is a strenuous task, and so any manual assignment of types is both error-prone and highly inefficient. In this thesis, we address the task of automatically assigning types to entities in a knowledge base.

Existing entity typing methods require great amounts of information about the knowledge base structure and properties, or assume that entity definitions are of a fixed nature. What we propose instead are two neural network architectures, one shallow and one deeper, which take short entity descriptions and, optionally, entity relationships as input. The goal to support knowledge bases with accurate entity typing of both existing and emerging types.

We experiment using the DBpedia knowledge base for evaluation, using two datasets: one reflecting accuracy on typing existing entities, the other on emerging entities. Results show that both our approaches are able to substantially and significantly outperform a state-of-the-art baseline, proving that neural networks can be used to support knowledge bases staying up-to-date and reduce overall incompleteness.