Siri or Watson?
I’ve been writing a story where I want to talk about the guts of Artificial Intelligence. This has often been thought to be 1) an expert system, or 2) a lexical inference system. The idea that there are two approaches is my own invention. As far as I know, it is not a distinction made at the MIT AI labs. I am using it in a fictional short story, so I am allowed to get away with it.
The first is a database approach where all the possible questions and answers are put into a database. A lexical analysis front end figures out what the question is, and it simply looks up the answer. I call this the Siri approach because, as far as I can tell, this is the way the Siri app works on the iPhone. The question is parsed and and analyzed, and the answer is retrieved. Somewhere in the nets is a giant Siri Database that is constantly being updated. The system (which I have never used) must have some way of getting feedback to see if it answered right. The feedback is then incorporated into the database look-up rules. The data that Siri returns is dynamic, in that it can get things not in its database like the local weather. It can also interact with the phone itself. Some of the interactions can be commands to control other apps such as dial the phone.
Siri uses lexical analysis in order to figure out what question you are asking, but I consider a Lexical Analysis engine to be another animal. Siri is more rigid in its responses because is uses a database of preloaded questions and answers. It is not intended to do things like write a Haiku about blackbirds in the snow (unless some joker has programmed it in.)
The term I use in my story for massive Lexical Analysis is a “Watson” after the IBM machine that won at Jeopardy. A Watson engine will have a database of questions and answers like Siri, but it will also perform a more extensive lexical analysis that will place the question in context. It will be able to search its database for context as well as simple question answer pairs. It will be able to answer questions that it has never seen before by finding information related to the context of a question, in addition to the simple words of the question.
Watson might not be able to write a Haiku, but should be able to recognize one. It does more than just retrieve data, it makes value judgements about the data that it retrieves.
Watson flubbed the answer to “Its largest airport was named for a World War II hero; its second largest, for a World War II battle.” It understood airports, WWII heroes and battles, but it did not seem to have the ability to create the context where all of these true. The answer is Chicago, which Watson had as its second choice, but it chose Toronto instead. It had trouble connecting Airports and Cities, with the answers. Since the question did not ask for a city, it wasted all of its time thinking about airports and couldn’t find much connection to WWII in the time allotted. It undervalued city in the context of the question and over valued airport. It spent its time looking at airports for WWII references and did not connect the results correctly to a city.
You can see the difference in an Expert System like Siri and real lexical analysis like Watson uses. At the base of each is a huge database, but Watson creates a context for searching out an answer.
In the future, computers will use lexical front ends like Siri or Watson. Some will be simplified, and some will be deep. Perhaps a Watson-like computer will constantly fine tune a Siri-like database, correcting and updating the questions and answers. Perhaps there will be a two stage analysis. Simple questions answered by wrote off the standard list of questions and answers, and another Watson-like layer that will work on the really hard questions that it has never seen before.
The question is really, which method is more like the way humans work? How do we answer questions? We try to figure out questions to things we don’t know, and try to remember the answer to questions we already know. We have Siri sense and Watson sense. I think, though, that a human is much more aware of the context of a question and more easily finds a framework for an information search based on the situation. Humans are more concerned with how a appropriate the answer is for the context of the question. Answering a question on an SAT test is much different from answering the question in a conversation with the postman. Even in the SAT test, it depends on whether a question is asked in a literature question or an historical question.
So, in my story, I have think about the underlying engine that allows a system to act human. Is it Siri? Is it Watson? Is it something else? I tend to think that there are monitor systems, that track the situation, subtext, history and plans ahead to make sure the underlying intelligence engine gets the right questions, hints and then filters the possible answers as to what is appropriate to the context. I might begin to think that a system to keep track of the context might be much more complicated than the knowledge base itself.
The last part of the question my story asks is at what point have we left the realm of creating a Turing machine that is indistinguishable from a human, to creating an actual human self-aware intelligence?
My story is basically an analysis of the nature of love, which suddenly makes the whole discussion a little more interesting. Can a human be fooled into loving a machine that is so like a human that a person can’t tell it’s not? Can a machine be so completely programmed that it can fall in love, even though it is only a response generated by the determinate functions of a complex set of algorithms. What is the difference? Does the difference matter? By asking these questions I hope that the story will help look at love in different ways.
This is weighty stuff and it is hard to write about without putting my readers to sleep. That’s why I am doing some of the heavier lifting here. (Wake up now and go back to work).
