Q: Since Watson’s successful debut on Jeopardy, there’s been a suggestion that computers can do natural language processing to improve customer service. What do you think?
Robert: I think it’s fascinating, and promising to imagine a computer can handle and process customer contacts, and be virtually indistinguishable from a human (The Turing Test).
But we’re a long way away from using machines to IMPROVE service from the customer’s point of view, never mind the costs, and resources required to run super computers that can do free form natural language WELL enough.
Q: It seemed like Watson did well, though. How accurate does a machine have to be to pass the Turing test?
Robert: Let’s keep in mind that we want a system where the computer doesn’t ever have to say “I don’t understand that”, or “Please repeat that”, or even “Did you mean…”. That’s hugely annoying to customers.
Q: Ok. I get that.
Robert: Good. So let’s say we have a machine that is 95% accurate in parsing natural language, and responding appropriately. That sounds pretty good, right?
Q: 95% would be amazing considering accents, language differences, and the huge range of possible customer requests and comments, so yes, that would be good.
Robert: Ok, so let’s say an average customer interaction conversation with a human involves, 500 words — that is the customer speaks that many, not all at once but in the conversation.
If the computer is successful with 95% of the words, then you will have 25 actual errors in the conversation. That’s HUGE. Imagine a conversation where one party doesn’t understand correctly 25 words, and has to stop and ask about the meaning? It doesn’t work, even at 95% accuracy.
Q: When you put it that way, it’s hard to imagine a machine handling open type interactions, at least now.
A: And don’t forget that we’re talking about more than the kind of voice recognition that one can do on one’s own computer, or even on a smart phone. It’s not just the computer figuring out the words or numbers, but the MEANING, and then providing the proper response.
Q: So you are saying that it will be a long time yet before we can use machines to process customer queries?
Robert: Yes, a long time, BUT what we can do now, albeit not well enough, is to have voice input that has constraints. For example, a computer can hear your bank account number over the phone, and do pretty well at making sense of it, or it have you choose from a limited menu of options.
That isn’t natural language processing though.
The bottom line: Technology has to be at least 99% accurate and successful if it is to improve service, so one can dispense with human backup, and not completely frustrate customers.
Got a story about voice recognition? Something that went right? Or really wrong? Add your thoughts in the comment section.