MSc Thesis Defense by Sune Andreas Dybro Debel


Deep Multi-Task Learning For Relation Extraction


In this thesis we investigate the usefulness of a multi-task, convolutional neural network ar- chitecture for relation classification. We review the relevant theoretical and practical research literature for relation classification, supervised machine learning, convolutional neural net- works, and deep multi-task learning. We test a state-of-the-art multi-task convolutional neural network architecture designed for relation classification on the SemEval 2010 Task 8 dataset using several hard weight sharing strategies. We investigate the sample complexity dynamics of learning SemEval 2010 Task 8 simultaneously with other natural language processing tasks. We find that only one of the proposed weight sharing strategies lead to improvements in generalization performance of this target task. We identify potential causes for this difference in generalization performance across weight sharing strategies, and make recommendations for further experimentation in this direction.

Supervisor: Dirk Hovy

Censor: Andrea Corradini