Do Boats and Ocean Suggest Beach? Case Studies of Pragmatics in Neural Dialogue Description Models

MPhil Thesis Defence


Title: "Do Boats and Ocean Suggest Beach? Case Studies of Pragmatics in 
Neural Dialogue Description Models"

By

Mr. Haojie PAN



Abstract

Pragmatics studies what a language expression really conveys in our 
natural languages. In this paper, we use a dialogue description generation 
task to study whether missing contexts and knowledge are learned and 
inferred. We use the aligned visual dialogue and image captioning datasets 
to learn how to generate a dialogue description based on textual dialogue 
turns. We also remove some dialogue turns to test the robustness of both 
learning and inference of a neural dialogue description model. To evaluate 
results, we use both traditional natural language generation related 
automatic metrics and our developed metric based on out-of-context 
concepts to compare typical neural network based models and a dialogue 
interaction-based model. We show that learning models can incorporate the 
context and some background or world knowledge learned from the existing 
large training corpus into the dialogue description.


Date:			Monday, 6 January 2020

Time:			2:30pm - 4:30pm

Venue:			Room 4475
 			Lifts 25/26

Committee Members:	Dr. Yangqiu Song (Supervisor)
 			Dr. Brian Mak (Chairperson)
 			Dr. Xiaojuan Ma


**** ALL are Welcome ****