On the integration of conceptual hierarchies with deep learning for explainable open-domain question answering

Tayyar Madabushi, Harish ORCID: 0000-0001-5260-3653 (2019). On the integration of conceptual hierarchies with deep learning for explainable open-domain question answering. University of Birmingham. Ph.D.

Preview

TayyarMadabushi2019PhD.pdf
Text - Redacted Version
Available under License All rights reserved.
Download (1MB) | Preview

Abstract

Question Answering, with its potential to make human-computer interactions more intuitive, has had a revival in recent years with the influx of deep learning methods into natural language processing and the simultaneous adoption of personal assistants such as Siri, Google Now, and Alexa. Unfortunately, Question Classification, an essential element of question answering, which classifies questions based on the class of the expected answer had been overlooked. Although the task of question classification was explicitly developed for use in question answering systems, the more advanced task of question classification, which classifies questions into between fifty and a hundred question classes, had developed into independent tasks with no application in question answering.

The work presented in this thesis bridges this gap by making use of fine-grained question classification for answer selection, arguably the most challenging subtask of question answering, and hence the defacto standard of measure of its performance on question answering. The use of question classification in a downstream task required significant improvement to question classification, which was achieved in this work by integrating linguistic information and deep learning through what we call Types, a novel method of representing Concepts.

Our work on a purely rule-based system for fine-grained Question Classification using Types achieved an accuracy of 97.2%, close to a 6 point improvement over the previous state of the art and has remained state of the art in question classification for over two years. The integration of these question classes and a deep learning model for Answer Selection resulted in MRR and MAP scores which outperform the current state of the art by between 3 and 5 points on both versions of a standard test set.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):