A User-Oriented Statistical Model for Word Prediction

Author: Madhuvanthi Muralidharan

Muralidharan, Madhuvanthi, 2017 A User-Oriented Statistical Model for Word Prediction, Flinders University, School of Computer Science, Engineering and Mathematics

This electronic version is made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.

Abstract

Word prediction has seen a vast range of applications over the years including SMS smart-messaging systems, and recently, in Assistive Technologies. The primary intention of word prediction is to reduce typing time, and while most prediction systems in Assistive Technologies achieve this goal, there is a requirement for the user to type using a keyboard. This is difficult for many disabled persons, but borders impossible for users suffering with severe speech or motor impairments. To enable such persons to type, a user interface is proposed in which the user is able to type by using predictions – at least from the users’ perspective – eliminating the requirement of a keyboard. This interface is part of a broader project on Unconscious Computer Interface where Brain Computer Interfaces are used to detect implicit intention rather than requiring explicit action. The aim of this project was to build a word-prediction model for the aforementioned user interface. Most word prediction models are based on statistical language models, in particular the n-gram model. Many improvements have been suggested to improve upon its performance, but little focus has been placed on user-based language models. Thus, the project aimed to address this research gap by building a user-oriented word-prediction model that is capable of generating predictions based largely on the user than the language. The model was designed as a back-off language adaptation model, drawing upon dynamic and static corpora to generate the predictions. Baseline studies validated the model as a word prediction tool, as well as identifying design elements with potential to improve performance. It is anticipated that the model will be integrated with the proposed Unconscious Computer Interface in future.

Keywords: Word Prediction, BCI, Statistical Model, User-Based, Quadriplegia, Locked-in-Syndrome
Subject: Computer Science thesis

Thesis type: Masters
Completed: 2017
School: School of Computer Science, Engineering and Mathematics
Supervisor: Professor David Powers