Miao Chen

Date of Award

8-2013

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science and Technology

Advisor(s)

Jian Qin

Keywords

automatic speech scoring, content scoring of speech, natural language processing, ontology, ontology-based representation, text mining

Subject Categories

Library and Information Science

Abstract

Text representation is a process of transforming text into some formats that computer systems can use for subsequent information-related tasks such as text classification. Representing text faces two main challenges: meaningfulness of representation and unknown terms. Research has shown evidence that these challenges can be resolved by using the rich semantics in ontologies. This study aims to address these challenges by using ontology-based representation and unknown term reasoning approaches in the context of content scoring of speech, which is a less explored area compared to some common ones such as categorizing text corpus (e.g. 20 newsgroups and Reuters).

From the perspective of language assessment, the increasing amount of language learners taking second language tests makes automatic scoring an attractive alternative to human scoring for delivering rapid and objective scores of written and spoken test responses. This study focuses on the speaking section of second language tests and investigates ontology-based approaches to speech scoring. Most previous automated speech scoring systems for spontaneous responses of test takers assess speech by primarily using acoustic features such as fluency and pronunciation, while text features are less involved and exploited. As content is an integral part of speech, the study is motivated by the lack of rich text features in speech scoring and is designed to examine the effects of different text features on scoring performance.

A central question to the study is how speech transcript content can be represented in an appropriate means for speech scoring. Previously used approaches from essay and speech scoring systems include bag-of-words and latent semantic analysis representations, which are adopted as baselines in this study; the experimental approaches are ontology-based, which can help improving meaningfulness of representation units and estimating importance of unknown terms. Two general domain ontologies, WordNet and Wikipedia, are used respectively for ontology-based representations. In addition to comparison between representation approaches, the author analyzes which parameter option leads to the best performance within a particular representation.

The experimental results show that on average, ontology-based representations slightly enhances speech scoring performance on all measurements when combined with the bag-of-words representation; reasoning of unknown terms can increase performance on one measurement (cos.w4) but decrease others. Due to the small data size, the significance test (t-test) shows that the enhancement of ontology-based representations is inconclusive.

The contributions of the study include: 1) it examines the effects of different representation approaches on speech scoring tasks; 2) it enhances the understanding of the mechanisms of representation approaches and their parameter options via in-depth analysis; 3) the representation methodology and framework can be applied to other tasks such as automatic essay scoring.

Access

Open Access

Download

Share

COinS