Natural language understanding and voice technology
In this episode The Future is Spoken podcast, Deborah Dahl, a natural language understanding expert, explores natural language understanding and its role in the world of voice technology.
What is natural language understanding? Natural-language understanding or natural-language interpretation is a subtopic of natural-language processing in artificial intelligence (AI). Natural language understanding deals with machine reading comprehension. Natural language understanding is considered an AI-hard problem.
However, there’s actually a lot going on behind the scenes when humans have conversations. Each conversation humans have is unique; the precise flow of words has never taken place before and will never take place again. Next, conversations and language are open-ended. This is why, for AI, natural language understanding is a real challenge, and classed as a ‘hard’ AI problem.
To be able to even come close to conversing the way humans do, AI will ultimately have to learn to process things that have never been said before and may never be said again. This is what fascinates Deborah. She has a Ph.D. in linguistics, and with her long interest in computers, she is passionate about computational linguistics.
Natural language understanding is a hot topic for conversational designers because they have to know what natural language understanding can and cannot do when they create a voice bot or interface, along with the current state of the technology.
Deborah gives some fascinating examples of some of the challenges conversational designers and natural language understanding experts face.
For example, it’s important for conversational designers to understand the concept of what AI experts call slots and form filling.
Explains Deborah: “If you ask a smart speaker to set an alarm for five, it implicitly has an understanding of five. But it's missing a slot. It's missing the AM/PM slot. So it needs to follow up with the user and make sure it pins that down, which time the user is thinking about.”
Another issue is the way it’s easy for humans to introduce other topics into a conversation. For AI, though, incorporating a new topic is difficult, if not impossible. In AI, a topic is called a domain, and in AI, entering a new topic causes confusion. For example, if you are talking about golf with a friend, and then discuss golf balls, their colors, and cost, this is easy for humans, but difficult for AI.
Next, Deborah explains that multi-intent utterances are difficult for natural language understanding.
She says: “If you had a human personal assistant, you might say something like: ‘Can you find out if there's a nearby Thai restaurant and if there is, make reservations for four at eight o'clock?’. Here, you have what we call two intents - to find a restaurant and make reservations. For a lot of technical reasons, a task like this is really hard (for AI).”
She observes that things that we can accomplish today with AI are not due to science, but due to having faster computers.
“It’s a synergistic cycle - the computers get better and then the technology catches up, and then the computers get better and faster, and so it goes,” she explains.
In addition, speech recognition is progressing at a steady pace. In the 90s, speech recognition was so bad, a lot of the natural language design processes were aimed at correcting speech processing errors. Today, understanding speech is much better.
In this episode, Deborah also touches on the emotional angle of having a voice interface friend that doesn't irritate us and just listens to us. The demand for artificial friends is increasing. She noted that humans love to have artificial friends, particularly given that loneliness is endemic.
Find Deborah Dahl on Linkedin.