SEMINARS

TALP Talk: Machine Learning, Deep Learning and Reinforcement Learning for Query-Based Summarisation


Next Tuesday, Diego Molla-Aliod will present his work on Machine Learning, Deep Learning and Reinforcement Learning for Query-Based Summarisation
Tuesday  Oct 10th, at 12.00.
Room omega: s2-s205 (UPC, Campus Nord)

 

In this talk I will present my contribution to the BioASQ Challenge 5b, phase B. Participants in this challenge need to provide ideal answers to medical questions given relevant text. In my  contribution I explore the use of query-based summarisation techniques in order to generate the ideal answer. My experiments contrast the use of deep learning (Word embeddings, ConvNet, LSTM) and traditional approaches to summarisation. The results indicate that "first n" is a very strong baseline, and that machine learning methods do not improve on traditional methods. In the second part of the talk I will explore an approach to improve the machine learning results by means of reinforcement learning. Current machine learning approaches to text summarisation normally attempt to minimise the prediction error of individual text fragments, and the final summary is obtained by collating text fragments in a subsequent stage. By applying reinforcement learning it becomes possible to attempt to maximise the final summary evaluation score rather than minimise prediction error of individual text fragments. My experiments show that it is possible to train a policy gradient system using training data and learn a global policy that improves the results of a separate evaluation set.

BIO: Diego Molla-Aliod is a Senior Lecturer at Macquarie University, Sydney, Australia. His research focuses on text-based question answering, text information extraction, and text summarisation. He is a founding member of the Centre for Language Technology (CLT) research group at the Department of Computing of Macquarie University, the longest-established and largest Language Technology research group in Australia. He is also a founding member of ALTA (Australasian Language Technology Association).

Additional information