“I’ve seen this movie before and it doesn’t end well,” mutters a VC whose start-up is short of cash.
“We’ll always have Paris,” Rick tells Ilsa near the end of the movie “Casablanca.”
Both these scenarios illustrate a favorite point of David Waltz, the natural-language expert whom I met while he was senior scientist at Thinking Machines: “Words are not in themselves carriers of meaning, but serve merely as pointers to shared experience.” …or something like that! The meaning should come clear from my examples.
In each case, there’s something complex and familiar that both parties recognize – something well beyond the capacity of words to represent without a sentient, intelligent being to condense them into a pointer at one end and to revive the words into meaning at the other.
When that doesn’t happen, you get scenarios like these: “Let’s have lunch sometime,” says Mr. Big Shot.
“Yes, that would be great!” says Little Worm. “Next Tuesday?”
“Ermm, actually, I’m quite busy next week… In fact, I’m tied up the rest of the month.”
Or “I’d like a red dress that flatters my figure,” says the shopper. She looks at the billowing red tent the saleswoman produces and says, “That is not what I meant. That is not it, at all.”
Natural language rules?
This is all to set up a series of posts on a current fascination of mine, pattern recognition. Pattern recognition means that you recognize a common pattern in a variety of instances…and that you can also produce instances to illustrate the pattern (which of course is exactly what I am trying to do here – illustrate a general theme of pattern recognition with examples).
Most computer programs say “if A, then do B.” Pattern recognition helps you determine whether A is true.
Pattern recognition takes a variety of forms, from object recognition and facial recognition to natural-language processing, which might more aptly be called “meaning recognition.” Pattern recognition ranges from recognizing a person in a crowd (useful to certain government agencies) to recognizing who’s likely at fault in a dispute, who is probably committing fraud, whether Juan’s a good match for Alice next door, which people will like a certain movie, what pitch is most likely to land a new advertising account, who designed a particular dress.
The inputs range from images to descriptions of behavior and numerical data, to natural language. You can do a lot of pattern recognition just with statistics, but only if you have enough data – and outcomes – or models, to start with. (That’s partly why I’m so hopeful about pattern recognition; there is more data everywhere, from people’s buying habits to GPS records of their movements, sensor data about all the things we see, electronic medical records that someday will follow a few standard formats so we can match behavior, therapies, genomes and outcomes.)
Some people are good pattern-matchers without ever articulating what they do; some (yes) recognize and can explain exactly what they are doing. (That is, in tech-speak, some people work like a neural net, producing results from a black box, while others work like an expert system, following explicit rules.)
And other people can read pattern-describing self-help books till they are blue in the face and still not recognize the situations in which they should apply the advice. Consider this piece of advice, for example: “Don’t ever ask a prospect who has said no to change his mind. Just give him a new proposition that he can agree with.” That’s how pattern recognition by a good salesman – and non-recognition (by the prospect who agrees with something that restates what he rejected before) – can work in business.
But back to software. A couple of weeks ago Stefanie Olsen of News.com wrote a piece about new forms of search. “Google is not the end of history,” I (and evidently a few other people) told her. Nor is search the final application…
Using pattern recognition in a variety of situations is the next, very diverse, frontier. If natural-language search is – or will be – the fat front of natural-language recognition, a wide variety of applications will be its long tail. Let me run through some examples from which you can divine my meaning. (I’ll be posting more of these over the next few days as I get the details fact-checked.)
I’ll start with the beginning of current history, Google. Google indexes words and phrases, and then uses the presence of those words plus popularity (the number of webmasters’ links to a particular page) to determine the ranking of the results – a list of pages where the search terms appear. In fact, Google’s search algorithms do a little more than that – fooling around with synonyms, eliminating stop words, possibly noting some metadata (authors and dates, for example) and other undisclosed “tuning” – but it is concerned with words, not meanings. And all it indexes or analyzes is text on the Web; it knows nothing about anything that is not in words, on the Web.
The future lies in moving beyond both those constraints. One is going beyond the Web, into “real life” and other media, such as television and films (and advertising); more on that later.
The other is expanding search (and other capabilities) to the meanings of those words on the Web: that is, to concepts, story lines, relationships – verbs, not nouns. Time-Warner acquiring AOL, for example, is very different from AOL acquiring Time-Warner… yet Google could not distinguish between those two. What’s enticing but not yet widespread is the ability not just to find relevant content, but to put the content into more regular form: for example, to build a table showing ten acquisitions of parts manufacturers by vehicle makers, listing the acquiring company, the acquired company and the amount paid including both stock and securities.
Google can get you lots of relevant (and irrelevant) articles, but it can’t fill in such table. It can get you a list of movies with Jennifer Aniston in them, but only IMDB tk link (compiled by people, often working for studios eager to promote their movies) can tell you the ones in which Jennifer Aniston starred.
That could change if we get better at natural-language understanding. Reuters is already pretty good at the generate-a-table-of-acquisitions task. However, as I heard the story, it also wanted to produce such news stories automatically by extracting data from press releases, but the reporters objected and insist on writing these formulaic stories by themselves.
Aside from whatever Powerset (mentioned by Olsen on News.com; I am an investor and it is still in stealth) is doing, there are lots of companies using natural-language pattern recognition in ways well beyond plain old search. Often, it’s a two-way process: The user supplies some information, the system makes some educated guesses, and ultimately a situation is recognized with the user doing quality control by saying, “Yes, that is right.” Then the appropriate action can be taken – whether it’s to remedy a situation (a dispute or a disease, for example) or to take advantage of an opportunity (an undervalued stock or an attractive purchase).
Here’s one you may not have thought of this way (or heard of at all), but then, pattern recognition is my job. SquareTrade is in the business of resolving petty business disputes; its major partner is eBay, for whom it provides dispute-resolution services to eBay buyers and sellers. (It also does quality checks and offers “good-behavior” seals.) The system begins by asking the user to fill in a form with simple questions: What was the item? What was the problem? What are the amounts at stake? [Disclosure: I have a small investment in SquareTrade.]
SquareTrade uses a fairly limited vocabulary and resolves a fairly limited range of disputes. You could probably argue that it’s not AI at all. But that’s not really the point. The point is that it recognizes patterns: “This is the kind of problem where the product didn’t meet the buyer’s specs,” vs. “This is where it was broken during shipping,” or “This is one where the buyer changed his mind.” For each of these situations (As) there are fairly standard resolutions (Bs): “Send the product back and charge the seller two-way shipping costs,” “Fix the product and split the cost of repairs,” or “Send the product back, refund the money minus the cost of two-way shipping, but give the buyer a black mark for changing his mind.”
Imagine if we could do a better job of recognition, agreeing on the facts, and resolve more disputes with reference to a generally accepted set of rules. Imagine if SquareTrade could understand disputants’ free-form descriptions of the events rather than just have them fill in forms. And imagine if we could represent the body of law as an expert system, and then accurately recognize the situations and figure out which laws to apply. This is a bit of a digression; more on expert systems later.)
In another domain, using natural-language parsing without much regard to meaning, Vantage Laboratories uses recognition of grammar and usage to grade SAT essays.
Similar techniques apply in health care: recognizing patterns to diagnose diseases, and then applying standard treatments. The challenge here is accurate observation – and the kind of probing a good doctor can do to uncover problems the patient is unaware of or doesn’t want to discuss.
More, and more intelligent automation, has been the dream in health care and education for decades, and it is beginning to happen here and there. One factor is that traditionally you needed the doctor or nurse to observe things. Now, there are devices that hook up to the patient directly and can feed physical data into a monitoring device. Those systems will find only what they have been trained to look for, but that’s a good start.
Of course, a good doctor or a good teacher is always better than a machine, but if there are not enough good doctors or good teachers, then machines can help fill the gaps and handle the routine cases. But more on real-world recognition next time…
Coming up soon: Recognizing real-world objects and the business models that could support .
…and recognizing consumers’ buying patterns – and the objects they buy