Knowledge Engineering & NLP - NLP Modeling chat

Hi. folks! Welcome to this chat on Knowledge Engineering. Tonight I'm going to cover my perception of why Knowledge Engineering is a highly useful methodology for rapidly learning stronger modeling skills. Clarify your understanding of modeling & of why AI as a field is highly useful for us as NLPers!

OK, let's jump right in & get started! Anyone here have an American Express Card? No, I don't want to charge you (!) -- I'm asking for another reason! To those of you who do, have you ever tried to use your AMEX card somewhere unusual, and the authorization came back with a request to call in? I have! This happens a LOT with Amex, and increasingly so with other cards.

These are some good answers. "Artificial Intelligence" is the name applied to the pursuit of making "intelligent" machines/computers. And there are a LOT of different approaches to this. From the theory-ridden early days of AI, there emerged two primary approaches to AI. LEARNING SYSTEMS & DECISION SYSTEMS.

At the time, they had to take their model of human thinking and break it into pieces they could replicate with some level of accuracy (Only as accurate as their measurements were, of course). Now -- as to LEARNING systems what emerged were called NEURAL NETWORKS & later GENETIC ALGORITHMS. But the 1st real successes with AI -- FEASIBLE financially viable successes -- were with DECISION systems. Also known as EXPERT SYSTEMS. Now, here's where we BEGIN to START to draw parallels!

They're called Expert Systems because they are software systems designed to MAKE DECISIONS, just like human experts do/did/would do. Humans aren't machines, we don't ALWAYS do things the same way, but it is possible to write software to flexibly apply criteria against a solution in reproduceable sequences -- such that IF you can somehow ELICIT the KNOWLEDGE that an expert knows, then you can build the software to do the "right" thing in the right circumstances. Make sense? (y/n)

Who sees a connection yet between what I'm touching on -- and NLP Modeling? (that's rhetorical!) The problem for them at the time was to come up with software SHELLS that would allow a person to RAPIDLY code someone's knowledge without extended programming tasks. That was solved over time and there are some VERY sophisticated expert systems shells available today, across a WIDE spectrum of costs. Those costs are usually roughly relative to the extent to which you can get moment-to-moment support -- or NONE at all.

Now. Once you have the software, there's another problem -- you need to train someone to get INTO someone's head, so to speak, and PULL the knowledge out of the expert. Sometimes by tooth & nail! Here's a major dilemma: Experts don't have a freaking CLUE *how* they know what they know -- they just KNOW it! They spent years chunking up and its just IN there. And Knowledge Engineers (thats their title) are people trained to get in there & get it out, encode it, and build expert systems software. They have to know what the clues are for whats redundant, whats deleted/omitted. They have to be able to construct cases/situations that cause the expert to access the knowledge that they DON'T KNOW they have! Make sense still?

Briefly, a shell is an EMPTY expert system environment designed to make the entry of knowledge -- SIMPLE & FAST.

Now, just to give a little context, those that know a little about me know that I haven't been training all that long; about 2 years. Which doesn't seem like a lot of time. Not even to me! Before getting headlong into NLP -- I was a senior Knowledge Engineer for the world's most successful Expert Systems company. And I'd been an author/presenter at the American Association of AI conference, -- for writing an Innovative Application of AI.

Yes, could have. I give them that -- there were indeed people in my consulting org who could have done better or done it faster. No kidding. And also on the flip side, there's a huge difference between people who have solid Knowledge Engineering skills -- and those who never learned it well. Companies often try to build such systems on the cheap, from within. It almost always turns out to be a far far less expensive effort to pay top consulting rates to the people who have a consistent history of project delivery.

OK, let's back up. When you know how to elicit the kinds of decision making other people make -- (as I did back then, without any framework for NLP™ at the time) then you're good at doing something required for expert systems, called Knowledge Acquisition. Big term for something Very Simple! Modeling.

Now, this is not what in NLP™ is called Strategy Elicitation. Modeling is a mathematical logical process whereby you identify a calculus for a model of thinking or expertise. Such that the calculus or cohesive model includes all the major variables -- or qualities that change from one situation to another -- that adjust the outcomes. And then if that calculus is then applied in such a way as to predict outcomes effectively, or produce similar results again, consistently, then you've produced a model (of unknown quality, just from that description).

That is what must be done to engineer knowledge. Now. We're not yet onto learning systems; we're still just dealing with decision systems. But ask yourselves this question & let it sink in for a while: What if you could build into your brain a way of automatically optimizing the decision systems you build in your brain? Everyone with me? (y/n)

Moving right along, now, the major detriment to Expert Systems alone is that they represent a momentary snapshot of a big knowledge picture. They don't grow very well, given the software shells, and given the limitations of the ES model!

It is easily feasible to edit the knowledge... bit by bit... long term over a lot of time. Discretionary adjustments are usually fine. But speaking on the basis of financial viability, Knowledge Bases rarely survive major overhauls. They don't seem to operate well after several re-engineering phases; too many band-aids, perhaps. And often, after a few years, many knowledge bases do require such overhauls. Why? Because the companies embody & accumulate more business expertise & information, and their in-house experts & employees learn more. And they want their systems to grow with what they've learned.

All factual data is applied to chains of rules... to produce results. Now chunk back up! Learning Systems are amazing because we CAN indeed teach computers new things. We have neural networks, genetic algorithms... and other techniques for learning, for actually absorbing new knowledge.

The Army once built a neural network -- a software program with a virtual "black-box" brain, so to speak... that gets "trained" to recognize tanks in a field. They fed this neural network a 'training set' of photographs showing tanks in fields. The Neural Network said "yes there's a tank in this one.... no there's no tank in this one." Over time the net was trained to accurately recognize tanks in fields.

Neural Networks can not logically tell you HOW they learn, or HOW they know they have the right answer. When the neural nets are being trained, they're given sets of training data along with the right answers. So they learn that this picture gets this answer, and that picture gets that answer. They train themselves to provide responses that match the training data's expected results. They *tune* themselves to get the 'right' answers. Its still a highly valuable technique because all you have to do is feed in the training set and GRADE the responses that come out (& yes, that's a simplification). Again, the problem there is there's no way to find out how it learned.

Here's why that's problematic: The Army, all spit-shined & hot-to-trot, took its fancy new system out into the field and had it look at real tanks rolling through the hillside. And EVERY SINGLE time, it screwed up & got the wrong answer, Nope, no tank there... This obviously did not please the Army.

So the generals asked "so WHY isn't it working?" And the techs said, "we don't have a clue." And they were telling the truth; no one thought to recognize that every photo they fed to the system was of a SUNNY day... until later -- after they'd evaluated the problems for a while. It rained on the day they tested the system in the field. Now, they were lucky. Not every neural network gets that far. But its still a useful approach. You know how to learn VERY VERY WELL.... don't you? In at least some context? And you know how to rationally make decisions in some circumstances, VERY VERY WELL.... right? So... if one approach (expert systems) is incredible for making straight-forward rational decisions AND can tell you EXACTLY how it made up its mind about something (and it can, they're very good for that!)

AND

if another approach (e.g. neural networks) doesn't have a clue HOW it learns but its amazing at learning...

THEN...

Where do you think we want to look when we're exploring advancing HOW we optimize our internal processes? Do we even look at that? We *could* abandon ALL that -- and I think that may be a useful choice in many contexts. And we can also look very closely... and do what I think you NLPers out there have heard before... take what's useful & Discard what's not! So what if we took the best of the learning systems... and the most effective parts of decision-systems... and you fuse them together inside your mind (take a moment to do that now)! Then for example you might get something like this: A hybrid system where you optimize your results -- by distributing the tasks!

After that ONE investment house started dominating the market, it took other investment houses a few months to catch up. This caused what we might call a shortterm blip in the market. And that exemplifies AI evolving in the marketplace. But you don't have to look there. You can just look innnnnnnside. To find the most advanced REAL intelligence there is. Because just as we try to make computers think like people... we can try to make ourselves think like computers (in SELECT circumstances) and we can learn from the field which has done more for cognitive thought modeling than any other field around -- CERTAINLY more than psychology's done.

As NLPers... AND As Knowledge Engineers... WE LOOK AT HOW. Screw "why." So to step outside a bit -- I did that for businesses for a few years, and when I started getting turned on to NLP™ -- it was like *click*, *click* -- I get all this -- and in some cases I had a lot of nuance work to do -- and in some cases I just had to allow for new Names for things I was already doing well. That should explain a little of the rapidity with which I integrated NLP™ skills in that 1st year after hearing about it.

Fortunately I also learned that the more you learn, the more there is to learn, so the initial overconfidence was replaced by a more open mind and a willingness to keep exploring & pursuing mastery & having a great time getting results!

Now, in practice, when you explore some of the Knowledge Engineering skills -- for encoding knowledge... you learn how to elicit strategies effectively, and the more you do it, the closer you're getting to learning how to REALLY model behavior & knowledge. So -- in London -- I'm going to lead people through learning how to Acquire Knowledge through some KE techniques, encode that knowledge, and then apply optimizing filters to it to help get to the most concise & elegant representation of that. And then USE that encoded knowledge to handle new circumstances well. And to then find out how much more smoothly their recognition of NLP™ patterns will occur! Not to mention how much more easily they'll understand and USE complex syntax. Not to mention how much more easily their use of the MetaModel will flow.

Let's finish with opening up the floor to questions regarding the AI & KE & NLP covered tonight!

Floydd's question 1st! Would you clarify your question for the channel: "how does that evaluate the classical AI (computer program) account of the intentionality (aboutness) of people's beliefs?"

There is much more to learn about -- including what I'm not doing just here in a Chat, -AND- including exercises to actually do a bit of knowledge engineering, AND including a working innate & practical understanding of a few Logic & Theorem proving techniques, like... Modus Ponens... Modus Talens... & Resolution. At the same time --

This chat DOES cover from a birds eye view an enormous amount of the connections between the two systems... those who have some knowledge of both will find it highly enlightening... those who have some or more knowledge of one may find the other more useful to learn about. Lastly, I was asked to recommend some Courses & Books! For those who wish to read more, look over these books, listed in order of my preferences, if you're looking for good, easily accessible material on AI & Knowledge Engineering. These books range from ~ $40-90. (Some are textbooks but are still excellent!) Check out these titles through your favorite BIG computer bookstore, or, through Amazon.Com:

• "Expert Systems: Principles and Programming" by Giarratano & Riley
• "Artificial Intelligence" by Winston
• "Expert Systems: Design and Development" by Durkin & Durkin
• "The Essence of Artificial Intelligence" by Cawsey
• "Introduction to Expert Systems" by Jackson
• "The Handbook of Applied Expert Systems" by Liebowitz (Editor)
• "Decision Support Systems and Intelligent Systems" by Turban & Aronson
• "Intelligent Decision Support Methods: The Science of Knowledge Work" by Dhar & Stein
• "The Engineering of Knowledge-Based Systems: Theory and Practice" by Gonzalez & Dankel.