Abstract

With advancements in artificial intelligence (AI), several augmentative and alternative communication (AAC) devices have begun incorporating AI with the purpose of accelerating and improving message transmission for users. The results have shown that AI can significantly reduce a user's keystrokes, but the generated responses may lack personalization. The objective of this study was to investigate whether integrating a personalized profile containing user information enhances the perceived accuracy and personalization of AI generated messages. Various methods of collecting profile information and differing levels of information within the profile were evaluated. One hundred eighty-one neurotypical English-speakers with no history of language disorder completed a Qualtrics survey. Each participant was assigned to one of six profile types (extensive free-recall, concise free-recall, extensive guided-recall, concise guided-recall, extensive multiple-choice, or concise multiple-choice). The profile was then incorporated into Chat-GPT 4o to generate responses to conversational scenarios. The participants were asked to indicate their preference between profile-integrated responses and non-integrated responses, as well as rate on Likert scales the likelihood to use a profile-integrated response, the accuracy of the response, and the personalization of the response. Mixed effects logistical regression analysis and mixed effects ordinal regression analysis suggest that incorporating a personalized profile can lead to generated responses that are more preferred by users compared to non-integrated responses, due to greater message personalization and accuracy. However, the method of gathering profile information and the amount of information in that profile do influence the effectiveness of the generated responses. Results from this study indicated that concise and extensive guided-recall profiles--structured around specific conversational categories--are the most effective. Both extensive and concise multiple-choice options produced responses that were less preferred by users than non-profile-integrated responses. Free-recall was seen to be effective when enough information was provided, but often users reported that their profile did not capture enough of the right content to be effective. It is hoped that personalized profiles developed through these methods will be incorporated into future AI AAC research and eventually implemented in AAC devices.

Degree

MS

College and Department

David O. McKay School of Education; Communication Disorders

Rights

https://lib.byu.edu/about/copyright/

Date Submitted

2025-06-12

Document Type

Thesis

Handle

http://hdl.lib.byu.edu/1877/etd13718

Keywords

augmentative and alternative communication devices, artificial intelligence, language learning models, user profiles

Language

english

Download

Included in

Education Commons

Share

COinS