This Is AuburnElectronic Theses and Dissertations

Employing Machine Learning Techniques in Conjunction with Linguistic Scoring Mechanism to Formalize the Uncertainty Carried by Scoring Numeric Processes Driven by Probabilistic Methods

Date

2020-08-18

Author

Livio Zavarce, Javier

Type of Degree

PhD Dissertation

Department

Computer Science and Software Engineering

Restriction Status

EMBARGOED

Restriction Type

Auburn University Users

Date Available

08-19-2022

Abstract

The aim of this research is to integrate the naturalistic mechanisms inherent to domain experts’ heuristics with machine learning capabilities as a useful framework with the capacity to support several domains of the human decision-making process. Because uncertainty translates as an imprecision, a property of the phenomena itself as another attribute (i.e., dimension or variable), we look at traits (non-statistical uncertainty) besides quantities (statistical uncertainty), confirming that the amount of fuzziness is not connected in any way to the quantification of information, it is an important trait of information representation. We present the approximated solutions of two highly complex systems whose behavior is not well understood (i.e. beyond the organized simplicity perspective). This was achieved by modeling their uncertainty, utilizing both linguistic scoring techniques, statistical analyses, and machine learning models. In both solutions, several experiments were conducted, and the results indicate that two complex non-determinist systems can be unified, and that linguistic scoring assists the decision-making processes and improves these systems’ through machine learning models. The first research domain is the modeling of domain expertise in the evaluation of coffee roasting and the process of coffee cupping. We have developed advanced methods based on machine learning solutions to solve these complex computational problems and unify the experts' knowledge from these domains and integrate these two processes through a Fuzzy Controller to impact the final quality grade in specialty coffees. Our research fills the gap in knowledge between the coffee roasting process and its impact on the final quality grading score, we look for the unification of these two processes. The second domain investigated speech intelligibility and we developed a model for standardized measurement in this area. Young children present a clinical challenge in speech intelligibility estimation, and as speech becomes less intelligible, unfamiliar speech patterns become more difficult to assess even amongst highly-trained clinicians. We investigated whether or not Linguistic assessment methodologies improve intelligibility assessment, we look either for error or bias. Moreover, we have developed advanced methods based on deep learning solutions (beyond regression models) to solve this complex computational problem. This work supports speech intelligibility research in the pediatric population (expandable to the adult population) and develops deep learning models for automatic speech intelligibility detection that captures both the abnormal speech variation and subjective ratings of assessors of speech intelligibility.