IBM’s Watson computer, which defeated the two all-time human champs on the TV quiz show Jeopardy! in 2011, is a glowing example of the heights achievable by predictive analytics. This is a machine that answers questions—about any of a broad, open range of topics. The same core technology that companies use to predict whether you’ll buy and which ad you’ll click is employed under Watson’s hood to predict, given a question, whether a candidate answer is correct. With this capability in place, Watson can “cast a wide net” by collecting thousands of candidate answers for a question, and then narrow down to the correct answer by predicting for each, “Is this the right answer?”
But, given that many of us have Siri, the iPhone’s eager-to-please personal assistant, right in our pocket, what’s so special about IBM’s one-of-a-kind, multi-refrigerator-sized monstrosity that cost tens of millions of dollars to build? How do the two compare?
First introduced as the main selling point to distinguish the iPhone 4S from the preceding model, Siri responds to a broad, expanding range of voice commands and inquiries directed toward your iPhone.
Siri handles simpler language than Watson does: Users tailor requests for Siri knowing that they’re speaking to a computer, whereas Watson fields Jeopardy!’s clever, wordy, information-packed questions that have been written with only humans in mind, without regard or consideration for the possibility that a machine might be answering. Because of this, Siri’s underlying technology is designed to solve a different, simpler variant of the human language problem.
Although Siri responds to an impressively wide range of language usage, such that users can address the device in a casual manner with little or no prior instruction, people know that computers are rigid and will naturally constrain their inquiries. Someone might request, “Set an appointment for tomorrow at 2 o’clock for coffee with Bill,” but will probably not say, “Set an appointment with that guy I ate lunch with a lot last month who has a Yahoo! e-mail address,” and will definitely not say, “I want to find out when my tall, handsome friend from Wyoming feels like discussing our start-up idea in the next couple weeks.”
Siri flexibly handles relatively simple phrases that pertain to smartphone tasks such as placing calls, text messaging, performing Internet searches, and employing map and calendar functions (she’s your social techretary).
Siri also fields general questions, but it does not attempt full open question answering, as Watson does. Invoking a system called WolframAlpha (accessible for free online), it answers simply phrased, fact-based questions via database lookup; the system can only provide answers calculated from facts that appear explicitly in the structured, uniform tables of a database, such as:
The birthdates of famous people—How old was Elton John in 1976?
Astronomical facts—How long does it take light to go to the moon?
Geography—What is the biggest city in Texas?
Health care—What country has the highest average life expectancy?
One must phrase questions in a simple form, since WolframAlpha is designed first to compute answers from tables of data, and only secondarily to attempt to handle complicated grammar.
Siri processes spoken inquiries, whereas Watson processes transcribed questions. Researchers generally approach processing speech (speech recognition) as a separate problem from processing text. There is more room for error when a system attempts to transcribe spoken language before also interpreting it, as Siri does.
Siri includes a dictionary of humorous canned responses. If you ask Siri about its origin with, “Who’s your daddy?” it will respond, “I know this must mean something . . . everybody keeps asking me this question.” This should not be taken to imply adept human language processing.
Siri and WolframAlpha’s question answering performance is continually improved by ongoing research and development efforts, guided in part by the constant flow of incoming user queries.
For more information on Watson’s impressive achievements answering human questions — and my thoughts on what makes it intelligent — see this article on Big Think.