4:24 p.m., Oct. 28, 2008----Keith Trnka, a doctoral candidate in the University of Delaware's Department of Computer and Information Sciences, won the Student Research Competition at Assets 2008, the 10th ACM Conference on Computers and Accessibility. The conference was held in Nova Scotia from Oct. 13-15.
ACM, the Association for Computing Machinery, has a number of special interest groups; one of those groups, SIGACCESS, promotes the interests of professionals working on research and development of computing and information technology to help persons with disabilities.
To win, Trnka had to advance through a multi-level competition, beginning with submission of a paper and then a poster. Finalists then delivered a 15-minute presentation on their work. Trnka emerged as the overall winner.
His paper, “Adapting Word Prediction to Subject Matter Without Topic-labeled Data,” presents an improved method for predicting words in augmentative and alternative communication (AAC) systems.
Trnka explains that such systems help people with speech impairments to communicate by speaking words for them. However, many AAC users also have reduced motor control, which slows down communication with AAC devices. Word prediction reduces the effort of producing text and increases the communication rate by predicting the desired words and allowing the user to select them using fewer keystrokes.
AAC devices are used for a variety of purposes--for example, e-mail, casual conversation, school assignments and recreation--and models are typically tailored to one of these topics. Consequently, a model that works well in one setting may not be as effective with another topic.
Trnka has developed a method, called fine-grained topic modeling, that treats each document as a topic, eliminating the need to split a corpus, or collection of text, into topics. Results obtained using the new method are comparable to those resulting when human-labeled topics are used.
“What sets Keith's work apart is that he has developed novel ways to zero in on those aspects of the corpus that are most like what the person has typed so far,” says his adviser, Kathleen McCoy, professor of both linguistic and cognitive science and of computer and information sciences. “This has enabled his algorithms to work better than what has been done before. His work could make a big difference in the communication rate of people who use these devices.”
Trnka, who earned his undergraduate degree at the College of New Jersey, expects to complete his doctoral degree in 2009. After graduation, he says he would like to continue doing research in his general area of interest, natural language programming, in either an academic or industrial setting.
Article by Diane Kukich
Photo by Ambre Alexander