Even though speech recognition is one of Yang Liu’s research interests, she confesses that there are times when, stuck on the phone in some automated customer-service system’s endless loop, she longs to talk with a human instead of a computer.
Still, she knows speech-recognition systems are here to stay in no small measure because researchers like her are quickly making them more sophisticated and flexible. “When I hear people complain about an automated-dialog system, I try to think about the technical problems that might have caused their experience to be bad. I’m happy to see that the speech technology has many users out there, but there is still room to improve the systems.”
An assistant professor of computer science at the University of Texas at Dallas, Liu works at the intersection of language and computing. Educated in China and at Purdue University, Liu found her academic calling early on. “I fell in love with speech and language processing as an undergraduate, and I’ve been working on it ever since,” she says. “It combines theory with real-world applications, and you can work on natural-language problems from the perspectives of either electrical engineering or computer science.”
Human language technology involves teaching computers to recognize human speech, turn information into normal-sounding human language, and represent human language in ways more easily used by computers. Liu has poured a lot of her young career’s research energy into finding ways for computers to accurately decode casual speech, such as the way people speak in meetings. “Our informal speech is chaotic,” Liu explains. “We make grammatical mistakes, stop in the middle of words, and throw in words and phrases we call discourse markers, like ‘like,’ ‘I mean,’ and ‘you know.’ Our applications can’t yet tell when different people speak, when people speak at the same time, or how people punctuate their spoken sentences.”
“We were interested in things like sentence boundaries, different types of utterances, the fillers people use when they pause, and disfluencies – the different ways we interrupt ourselves"
Enabling natural-language applications to more accurately decode and transcribe messy human speech is a major objective that many investigators in Liu’s field are working on. Liu decided to tackle one aspect of the problem in her Purdue dissertation (part of her dissertation was conducted while she was at the International Computer Science Institute in Berkeley). Her research dealt with the problem of automatically detecting the structural events in human speech. “We were interested in things like sentence boundaries, different types of utterances, the fillers people use when they pause, and disfluencies – the different ways we interrupt ourselves.”
Liu’s doctoral research focused on improving the automatic detection of these kinds of structural events so that automatically generated transcriptions of casual speech are easier for subsequent language-processing applications to use. In addition to speech recognition and understanding, Yang Liu works on developing effective natural language processing techniques that can help people cope with information overload. One of these projects is summarization – the automatic generation of summaries of spoken language, such as meetings. Liu says that her background in both electrical engineering and computer science is especially valuable for this and other speech and language-processing research.
Although she didn’t necessarily have her heart set on academic life, Yang Liu says she always knew she wanted to do research. Now that she’s here, she has surprised herself by how much she enjoys teaching. “When I first started teaching, I spent a lot of time on the preparation, and I was constantly thinking about the class and could not do other things in the morning if I had a class to teach in the afternoon.” Now completely at ease in the classroom, Liu says she is eager to refine her teaching skills by paying close attention to her students’ evaluations.
Having just joined the UT Dallas faculty in 2005, she is supervising three graduate students and one postdoc. She especially enjoys working with her students one-on-one. “I’m just learning to patiently take pleasure in my graduate students’ progress,” she says.