The Semantic Advantage

February 24, 2011

What’s up with Watson?

Filed under: products for semantic approach,semantic technology,Watson — Phil Murray @ 1:13 pm

The IBM/Watson Jeopardy! “challenge ” — three days of Jeopardy! matches first aired on Feb 14-16, 2011 — is a watershed event in the related worlds of Knowledge Representation (KR) and search technology. The match features IBM hardware, software, and data resources developed over seven years by a dedicated IBM team matching wits with two all-time Jeopardy! contestants. Mainstream media are playing it up, too. (Get the IBM perspective on Watson at http://www-943.ibm.com/innovation/us/watson/.)

The result: A big win for Watson. And IBM. And potentially very big losses for those working in the fields associated with Knowledge Representation and information search.

The angst of the KR community is evident in the posts to the Ontolog forum immediately preceding and during the televised challenge. (See the forum archives at http://ontolog.cim3.net/forum/ontolog-forum/.) for Feb. 9, 2011 and following days.) A profession already in “We need to make a better case for our profession.” mode received a  major jolt from IBM’s tour-de-force demonstration of “human” skills on a popular game show.

Although Watson incorporates significant ideas from the KR and search communities — it was, after all, developed by experts from those communities — it’s the effectiveness of the statistical component that drives much of the uneasiness of the KR community. Watson relies heavily on such statistical search techniques as the co-occurrence of words in texts. Lots of texts.

By contrast, the KR community focuses more heavily on interpreting and representing the meaning of natural language — usually by building a model of language from the ground up: concepts assembled according syntax. The results range from simple “taxonomies” that support advanced search in organizations to very large “computer ontologies” that can respond to open-ended natural-language queries and attempt to emulate human problem-solving. But none, so far, that can lay claim to besting smart humans in a challenge most think of as uniquely human.

So major sales of new search engines in big business are going to come to a screeching halt until upper management figures out what happened. All they know now is that an IBM machine outperformed two really smart humans in the domain of common knowledge and made their current and planned investments in search technology look like losing bets. Budget-chomping losers at that.

Why Watson?

Did IBM invest substantial expertise and millions of dollars of computer hardware and software to create what one contributor to the Ontolog forum called a “toy.” Yes, it is a “toy” in the sense that it is designed to play a quiz show.

But oh what an impressive toy! And you know it’s an important toy precisely because the people who understand it best — the members of the KR community — are really queasy about it, devoting hundreds of posts — many of them very defensive — to this subject on the Ontolog forum. Ever notice how participants in live political debates get louder and interrupt more frequently when the weaknesses in their arguments are being exposed?

The good news is that these discussions have surfaced and explored the root goals and benefits of the KR field itself — often in langauge that makes those goals and benefits more accessible to the outside world than discussions on the fine points of semantic theory.

IBM’s end game, of course, is quite simple:

  1. Demonstrate that the path it took has been successful — especially relative to other solutions —  and
  2. Make the buying public aware of that success.

And what could be a more perfect audience than diehard Jeopardy! watchers — millions of college-educated viewers every night, many of whom will influence buying decisions in business and government organizations. IBM consultants won’t have to explain what they’re talking about to non-technical decision makers. The decision makers will include more than a few Jeopardy! watchers. Even better, the mainstream media has been talking about the Watson challenge for days already, often misunderstanding and exagerratng the nature of Watson’s victory.

Score a big win for IBM. A really big win.

What does Watson do?

If you haven’t watched the three-day Jeopardy! event, you can find it in several places online. Beware of sites that charge for downloads.

The DeepQA/Watson project team leader, David Ferrucci, gives a very good explanation of how it works here: http://www-943.ibm.com/innovation/us/watson/watson-for-a-smarter-planet/building-a-jeopardy-champion/how-watson-works.html.

What Watson does not do

Watson is a brilliant achievement, both in terms of technology and marketing. But you need to take it all with a grain of salt. To begin with, the Jeopardy! categories chosen for this challenge have at least two significant constraints: No audio clues and no visual clues. Watson cannot “see” pictures or videos, and it responds only to electronically encoded text.

In theory, at least, those limitations could be overcome quite easily. We already have smartphone apps that will “listen” to a radio tune and tell you the name of that tune. Speech-recognition apps for smartphones and personal computers are remarkably good. Identifying the voice of a particular person seems plausible, too, if the detective shows are accurate. Facial recognition software and applications that identify other objects in static images are available now.

I’m not qualified to tell you how effective such applications are, but they seem impressive to me. And — just as Watson has extracted information from millions of texts for use during the show, there’s no reason to assume that its designers could not build structured descriptions of non-text resources prior to the show. Watson might, in fact, have a huge advantage in establishing matches with such non-text objects relative to humans … at least some of the time.

How the Jeopardy format is an advantage to Watson

The Jeopardy! format itself imposes inherent constraints — most of which are advantageous to the Watson team. And the IBM Watson team fully understands that. They just don’t talk about it too much — perhaps because what it does do is so remarkable.

  1. The Jeopardy clue team consciously limits the difficulty of each clue in several ways.
    • Some clues are harder than others, but most rely on “general knowledge.” Using its human experience, the clue team avoids clues that would be too difficult for the average smart person. Such constraints limit Watson’s advantage. Giving the value of pi to 10 places or listing all vice presidents of the US would be child’s play for Watson. When it comes to raw memory, Watson is going to win.
    • The clues rarely require analysis of complex conditions. After all, the object of the game is for humans to come up with the right question in a few seconds. The absence of more complex and subtle clues is generally an advantage for Watson.
    • The clues and questions fall within the cultural experience of Americans with a typical college education. Listing great Bollywood films would be easy for Watson but tough for most Americans. (That may change over time.)
  2. The response to most clues is a question that identifies a small set of concepts or entities — usually only one.
    • By “entity” I mean specific people, places, or things. [Who/What is] Henry VIII, Lake Nicaragua, and The Declaration of Independence are among the specific “questions” I have heard.
    • By “concept” I mean a class of things, whether concrete or abstract — like dogs, weaponry, human health, or happiness. I believe that if we took a statistical survey of Jeopardy! questions (the responses), we would find that the clue frequently consists of lists of things belonging to a class (definition by extension — a subset of the things in that class) rather than definition by intension (a set of properties that define a class). I suspect that this also favors Watson in a substantial way.

So Ken Jennings and Brad Rutter took a thumping on national television because categories that might have favored humans at this time were eliminated, and because there are other significant constraints imposed by the “rules” of the game itself. The thumping could have been worse. And IBM knew that.

So is Watson besting humans at a human skill?

In its Jeopardy! challenge, is Watson besting humans at a human skill? That’s the picture often painted in the media:

IBM trumpets Watson as a machine that can rival a human’s ability to answer questions posed in natural human language.

Source: Computer finishes off human opponents on ‘Jeopardy!’ By Jason Hanna, CNN
February 17, 2011

Well, it really depends on what you mean by “answering questions.” Sometimes you are looking for the name of a British monarch or slight changes in spelling that result in strange changes in meaning.

However, in most senses, what Watson’s designers have asked it to do is very simple when compared to what humans do when they answer questions. (See above, “How the Jeopardy format is an advantage to Watson.”)  Humans also do not ask random questions. (OK, your young children and some of your adult friends may do that, but those are different challenges.) In fact, your objective in asking a question is usually to carefully identify and frame the right question so that you improve your chances to get the answers you want … in order to address a specific problem. Unless, of course, you are a quiz-show contestant or taking a fill-in-the-blanks final history exam.

Keep in mind that, as more than one contributor to the Ontolog forum has observed, Watson doesn’t “understand” its responses. It only knows that its responses are correct when Alex Trebek says so. And, unlike in most human exchanges of meaning, it has no goals or purposes in mind, so it doesn’t know what the next question should be.

In many senses, Watson is an advanced search engine — like Google. Once you understand the nature of the game, there’s a temptation to call the Jeopardy!/Watson match a cheap parlor trick. But it wasn’t so cheap, was it? Still, brilliant work by the Watson team. Clever, too. (That’s not a criticism.) They really understood the nature of the game.

Watson got an unexpected boost from Alex Trebek, too, as Doug Foxvog noted on the Ontolog forum. My wife and I are longtime Jeopardy! watchers. It seems to us that Alex and his “clue team” have become increasingly arbitrary in their acceptance of specific answers, whether for the correct phrasing of the question or for error in facts. Some of their judgments are clearly wrong. That’s understandable. It’s the trend that irritates us, so we end up yelling at Alex. I guess we need to “get a life.”

Those are my abstract complaints. Looking at the multiple responses considered by Watson (shown on the bottom of the screen in the broadcast) gives you a gut feel for how little true “understanding” is involved. And you can be certain that the [types of] clues Watson responds to correctly are different from the types of clues humans respond to correctly. Statistically, there will be variance in the specific correct answers.

There’s more to be learned (by the general public, like me) about what actually happened by more careful analysis of the Jeopardy!/Watson challenge. But we need to let it go as a metaphor for computers outsmarting people.

Could Watson-like technology solve business problems?

Could Watson-like technology solve business problems? In some important ways, Yes. It could be customized to answer a variety of business-oriented questions with a high degree of confidence … and tell you how confident it was about the responses it provided. Applied to a narrow domain rather than the open-ended domain of common knowledge (as on Jeopardy!), Watson-like technology should have a high degree of confidence in most of its responses when retrieving information from a massive resource, and like a typical search engine, it should be able to tell you where it found those answers.

That’s truly valuable, especially when the retrieval problem is well understood. It might even qualify as a good return on investment, in spite of Peter Brown’s comment on the Ontolog forum:

That’s because “artificial intelligence” is neither. It is neither artificial – it requires massive human brainpower to lay down the main lines of the processors’ brute force attacks to any problem; and It is not intelligent – and I seriously worry that such a failed dismal experiment of the last century now re-emerges with respectable “semantic web” clothing.

Source: Posting to the Ontolog forum by Peter Brown, Re: [ontolog-forum] IBM Watson’s Final Jeopardy error “explanation”17-feb-2011, 9:27 am.

It won’t be cheap, at least initially. But that’s not the real problem. Watson team leader David Ferrucci himself brings up the medical/diagnostic possibilities. And who has the most money today, after all???!!!!

In the end, however, neither Watson nor Google nor the inevitable Watson-like imitators will do what we need most. Nor will the work of the KR community when it focuses solely on machine interpretation of natural language. Not by themselves.

Watson-like technologies also risk becoming the end itself — the beast that must be fed — just like the many current information technologies they are likely to replace. It will be a great tragedy if the KR community, the search community, and the organizations and individuals they serve assume that Watson-like approaches are the primary solution to today’s information-driven business problems.

But Watson-like technologies are an important complement to what we need most. As well as a brilliant achievement and a watershed event in technology.

November 15, 2009

An old favorite plus a new favorite = solution

Filed under: knowledge management,products for semantic approach — Phil Murray @ 3:03 pm

No rants about search engines this week. Instead, praise for a terrific desktop search engine — dtSearch — and ABC Amber SeaMonkey Converter, one of many converters offered by Yernar Shambayev’s Process Text Group.

Online searches lead mostly to … more online searches instead of to reusable value. But we can’t live without them, and I have to admit that Google and Yahoo! are steadily improving the effectiveness of their products. However, sometimes our needs are more narrowly defined than locating something in all the world’s information.

When I’m building a network of knowledge using the approach I have designed, I need to know whether the idea or concept I want to add to the network is the same as — or similar to — other ideas or concepts already in the network. Let me stop for a minute and define idea as an observation about reality — the equivalent in meaning to a simple sentence in natural language. Contrast that with concept — the essential name of a thing, whether material or imagined. Concept appears to be the preferred terminology for practitioners who construct taxonomies (or facets), thesauri, and ontologies that organize such entities into larger structures. I won’t go into the fine points here.

I have not built a rich ontology of the concepts in the ad hoc spaces I discuss, and I haven’t found any affordable tools that allow me to look for similarities among ideas. So I resort to a very simple practice: I maintain a directory in which each idea and each concept occupies a separate file. The file contains the name of the concept or idea, explanations of those items, and text examples that contain instances of those items. A full text search of that directory using the new concept or idea as the query retrieves the search engine’s best guess at files that contain similarities with concepts and ideas already in the network of knowledge.

Or not. Because most search engines are primarily string-matching tools, and the files retrieved may not be what I want.

dtSearch is better than that. In addition to the features you might expect in a good or desktop enterprise search engine — including stemming, wild cards, fuzzy search, proximity search, and Boolean operators — you have the option of looking for files that contain synonyms based on Princeton’s WordNet — a kind of semantic network that anyone can use. So even if you can’t keep track of synonyms, the dtSearch tools will. You can add your own synonyms, too, within dtSearch.

Great stuff. Some consider the dtSearch interface dated, but I think it’s highly functional. Real easy to set up separate named indexes for different sets of directories, too. (Excuse me. I’m dating myself. We call them “folders” now, don’t we?)

I also use dtSearch for a variety of other search tasks — including finding emails from the thousands I have captured in SeaMonkey. Making those emails accessible in a reasonable (and, ideally, consistent) way has been virtually impossible. The native SeaMonkey search features — like those in other email clients I have encountered — are simply inadequate.

And even if those email search features were superb, they wouldn’t solve the problem, because SeaMonkey stores each mailbox as one big file. I do mean big for some of my mailboxes. So finding a huge file is almost meaningless. Big files will satisfy many queries unless you use proximity searches and other tricks, and even if one mailbox does contain the information you want, it may take a long time to find the right spot within that file. And you have to go through the same process if you want to execute that query again.

ABC Amber SeaMonkey Converter solves that problem by allowing me to split SeaMonkey mailboxes into separate HTML files. (I could use ABC Amber options to convert them to text or a couple dozen other output formats, but I prefer HTML for a variety of reasons.) When I use a dtSearch query against the directories containing those exported HTML emails, I get a highly relevant selection of small files — exactly what I want.

Very easy, too. When I ran ABC Amber the first time, it found the SeaMonkey mailboxes automatically. The emails in each folder were displayed in a list, and you can easily select as many or as few as you wish. Oh, and I should mention that ABC Amber promotional pages stress the ability of the converter to output a single, integrated HTML file from a mailbox. That’s a plus for many people, but not what I want.

I also tested the mailbox-to-TreePad converter. (You just click a different output option in ABC Amber.) The results were flawless and the TreePad outliner let me explore view the email content by date. Cool.

One caution: As of this writing, it appears that SeaMonkey has changed where it places email folders. So folders I created with SeaMonkey 2.0 — and any new email since the changeover — did not show up in the ABC Amber converter, but I was able to redirect the program to the new location using an ABC Amber option. I have advised the Process Text people about this.

UPDATE (16-nov-2009): The ProcessText people have already updated the converter. It now finds the SeaMonkey 2.0 mailboxes automatically. That was quick!

I’ve been using dtSearch for nearly a decade now. It’s still worth the money — about $200 for an individual license. Adding ABC Amber SeaMonkey Converter (about $20) to my set of tools will really make a difference.

Create a free website or blog at WordPress.com.