MULTI-TASK LEARNING FOR QUESTION ANSWERING

MPhil Thesis Defence


Title: "MULTI-TASK LEARNING FOR QUESTION ANSWERING"

By

Mr. Tao ZHONG


Abstract

Nowadays, chatbots, or dialogue systems, become quite popular and lots of 
companies invest large amounts of money on them. Chatbots can be divided into 
two categories, namely opendomain bots and task-oriented bots. The big 
challenge in open-domain chatbots is that the domain is not limited. As for 
task-oriented chatbots, they focus on a particular domain such as booking 
flight tickets, etc.

Question answering (QA) in dialogue can be treated as a single-turn 
conversation. Two approaches are applied to produce answers, namely, 
retrieval-based approach and generation-based approach.

Retrieval-based question answering(QA) aims to select an appropriate answer 
from a predefined repository of QA according to a user’s question. Pervious 
research usually employs one kind of discriminative model such as dual encoder 
based neural network to improve the performance of QA classification, commonly 
resulting in overfitting. To deal with the problem, we investigate multi-task 
learning(MTL) as a regularization for retrieval-based QA, jointly training main 
task and auxiliary tasks with shared representations for exploiting 
commonalities and differences. Our main task is a QA classification. And we 
design two auxiliary tasks in MTL: 1) learning sequence mapping of actual QA 
pairs via sequence to sequence learning and 2) RNN language model without 
relying on labeled data. Experimental results on Ubuntu Dialogue Corpus 
demonstrate the superiorities of our proposed MTL method over baseline systems.

Generation-based question answering (QA), which usually based on seq2seq model, 
generates answers from scratch. One problem with seq2seq model is that it will 
generate high-frequency and generic answers, due to maximizing log-likelihood 
objective function. We investigate multi-task learning paradigm which takes 
seq2seq model as the main task and the binary QA classification as the 
auxiliary task. The main task and the auxiliary task are learned jointly, 
improving generalization and making full use of classification labels as extra 
evidence to guide the answer generalization. Experimental results on both 
automatic evaluations and human annotations demonstrate the superiorities of 
our proposed MTL method over baselines.


Date:			Thursday, 27 July 2017

Time:			3:00pm - 5:00pm

Venue:			Room 2612A
 			Lifts 31/32

Committee Members:	Prof. Qiang Yang (Supervisor)
 			Dr. Yangqiu Song (Chairperson)
 			Dr. Qiong Luo


**** ALL are Welcome ****