IBM Watson has already changed our perception of what computers can do -- it beat the best Jeopardy! champions, and it's being used for medical diagnoses. But what sets Watson apart? What makes it different?
1. It Reads Unstructured Text
When you feed data into a computer, traditionally it has been highly structured -- think a table listing all the U.S. Presidents, with columns for when their terms started and ended. Watson can read that kind of data, sure. But it specializes in reading raw human writing, also known as "unstructured data." You can feed it the biography of a president, and it will pick apart every sentence to learn what facts are contained in there. It will figure out all sorts of information within that huge body of text, and it doesn't require humans to put it all into a structured format first.
This ability to take in unstructured data is a huge strength for Watson. It means that the system can take in new bodies of knowledge quickly. You want it to know about medicine? Feed it the text of every medical journal you can find. You want it to learn Bible trivia? Feed it the Bible.
As we produce lots of information in unstructured form (for example, this blog post!), Watson is ready to consume it and make sense of it. As a trivia junkie, I can't wait to ask Watson some questions of my own.
2. We Train It
In addition to just dumping text into Watson, humans actually train the system to understand what's most important and reliable within the text. For instance, Watson pulled in all of Wikipedia prior to its Jeopardy! appearance, and stored that data offline. But it also had a huge corpus of other knowledge. Humans can tell Watson to trust one source of information (say, a biography of Bob Dylan) more than another (say, his Wikipedia entry). That doesn't mean the system ignores the less-trustworthy data -- but it knows which source to trust if there are conflicting facts.
But going deeper, when we think about Watson as a computing platform, we don't actually program Watson for new applications, per se. Instead of programming the computer, we train the computer using new data and human understanding of a topic. For instance, as a doctor you might train Watson to prefer newer medical journals over older ones -- so that data from the 1800s is taken with a grain of salt.
This shift from programming to training is part of why IBM calls this effort "Cognitive Computing." In the future, we will rely less on rote calculation, and more on interaction and learning.
3. It Asks Clarifying Questions
When Watson handles a tricky question in its current applications (like health care), it comes back with a set of possible results -- but it's also able to ask clarifying questions. It's clever enough to know that with a bit more information, it would be able to rule out an answer, or increase confidence in one of the answers it's already offering.
In health care, this could take the form of ordering a medical test. Presented with a series of facts about a patient, Watson could effectively say, "If you run this blood test, I'll have more confidence in my answer, or you can rule out these diseases." That's a very unusual thing for a computer to do, because it requires the computer to understand both what it knows and what it doesn't know. Knowledge may be power, but knowledge of your limitations is a superpower.
4. It Handles Open-Domain Questions
Most Question Answering systems are programmed to deal with a defined set of question types -- meaning you can only answer certain kinds of questions, phrased in certain ways, in order to get a response. Apple's Siri is an example of a closed-domain system. If I ask Siri a question, it has to be one of those questions Siri has been pre-programmed to answer (that's why so often, Siri gets confused and just offers to Google it for me). It's great when it works, but if you ask something just slightly out of its domain, the system falls apart.
But Watson is different. Watson handles "open-domain" questions, meaning anything you can think of to ask it. It uses Natural Language Processing (NLP) techniques to pick apart the words you give it, in order to "understand" the actual question being asked, even if you ask it in unusual ways. It also handles questions on any topic, combing through all the data it has, looking for the subject you're asking about.
IBM actually published a very helpful FAQ about Watson and IBM's DeepQA Project, a foundational technology used by Watson in generating hypotheses. My favorite question from that FAQ is: Is this going to be like HAL in 2001: A Space Odyssey? The answer is instructive (and I've added emphasis below):
Not exactly. The computer on Star Trek is a more appropriate comparison. The fictional computer system may be viewed as an interactive dialog agent that could answer questions and provide precise information on any topic. A primary goal for DeepQA is to greatly improve information seeking tasks over natural language content but ultimately, we would like to see the underlying technology help make computers more effective at communicating in human terms. Watson uses the DeepQA technology to push the envelope in natural language processing and automatic question answering. A powerful and fluent conversational agent, like the Star Trek computer, is a driving vision for this work.
I'll take the Trek computer over HAL any day. One to beam up!
5. It Shows Its Work
When Watson answers a question, it goes through a bunch of work to get there. First, Watson has to parse what kind of question is being asked, and what kind of answer is being sought. Second, Watson builds a series of hypothetical answers -- building a huge volume of possibilities, even if they're wrong. Third, it tests these hypotheses using a variety of different techniques, mostly based on the quality of the evidence. Finally, it merges and scores the possible answers: using its own question-answering history, the past reliability of various sources, and other techniques, Watson chooses the top answers, and presents them to a person.
But what's transformational here is that the person may then dig in and examine the underlying reasons that Watson chose those answers. During Jeopardy! we just got to see the top answers and a confidence score, but in a less time-sensitive application (like in a doctor's office, or when evaluating a given investment), humans can look at the answers as well as the supporting evidence. Because of this, humans can apply their own experience and expertise to decide whether that evidence is reliable. It's also easy to see how the evidence itself points to new areas of research -- if Watson tells you a medical study gave it confidence that an answer is correct, a doctor might want to go and read the whole study to see what else is in there.
