Relation Extraction With Synthetic Explanations And Neural Network

Abstract

Relation Extraction, which is defined as the detection of existing relations between a pair of entities in a sentence, has received a large interest lately, including more recent work on using neural methods. Since neural systems need a large number of annotated sentences to build effective models, Distant Supervision has been a preferred choice for collecting training labeled data. However, recent published work has shown that, training classifiers via a small number of annotated data and some explanation of why a sentence expresses a relation performs as accurate as distant supervision methods working with a large number of annotated sentences. In this thesis, we show that we can generate synthetic explanations, based on a small number of trigger words, for each relation in a way that the resulting explanations achieve comparable accuracy to human produced explanations by training a neural classifier. Our system is evaluated on five relation extraction tasks with different entity types (person-person, person-location, etc.) and the results show that synthetic explanations can work as precise as human generated explanations for the task of relation extraction. The proposed system also has the ability to classify noisy data coming from distant supervision methods with a reasonable accuracy.