ENTITY RELATION EXTRACTION Exploring the Use of Coreference Resolution in a Distant Supervision Approach to Automated Relation Extraction

Abstract

This Master’s thesis describes the effect of coreference resolution on a distant supervision approach to automated relation extraction. Coreference resolution is used as an addition to an existing relation extraction method, described in Mintz (2009). The thesis gives a detailed analysis of a reimplementation of this existing method, and provides an insight in how coreference information influences the performance. In this distant supervision approach to relation extraction, the database Freebase is used for the annotation of relations in Wikipedia text. A classifier is then trained to learn these relations between entities in the text. The main advantage of this approach is that the data is automatically annotated. This prevents the expenses of manual annotation, and makes it possible to train a classifier on a very large dataset. Using coreference information is a way of increasing the amount of data available and the expectation is that this will improve the result of the relation extraction system. An automatic evaluation method and a manual analysis of the performance are described, providing a detailed insight in the system’s behaviour. The evaluation shows that including coreference information does not improve the precision and recall of the system. However, it does enable the relation extraction system to find more relation instances, which shows that is does have an advantage and a lot of potential in future research.