Knoesphere: Building Expert Systems With Encyclopedic Knowledge

I. Introduction to the Knoesphere

This paper describes an ongoing . research project whose goal is to represent a comprehensive corpus of real world knowledge (both the stze and scope _of The Encyclopedia Britannica[l]) in a knowledge base; t_.e., as a structured network of concepts, rather than as pteces of text. The purpose is to enable a user t<? browse through that knowledge in as effective and flextble_ a manner as possible. The project can b_e concetved as the combination of three methodologtes: _expert sys~ems [2] (hitherto narrow in scope), encyclope~tas and onltne data bases [1,3] (hitherto only hghtly cross-mdexed assem)Jlages of prose) and videogames (hitherto rarely educatiOnal). We now present a scenar!o of a session with _Knoesphere. We include it for motJvatton of the proJect, and. to illustrate how the various levels of modelling an_d tailonng and filtering can affect what the user expenences.

It is Autumn, 1995. Sitting in our living room, w_e connect our home computer to the relevant encyclop~dtc service and don a helmet and gloves. The helmet IS a lighter: faster version of (4], with separate images for each eye, speakers for each ear, and sensors to track eyt? _and head movements. The gloves enable our hand postttons to be monitored, and provide pressure feedback.

Displayed before our eyes is a. trio of 0-ree­ dimensJOnal arrays, labelled Axes, Gu1d_es, and Filters. Each array is filled with icons. We pomt at the A~es array, and it grows laq~er ·. as though w~ were _approachmg it. We now see that tts tcons symbohze vanous sorts of dimensions: ways of organizing knowledge: by location, by time, by degree of cettainty, by principal objects involved, by philosophical orientation, by academic discipline, etc. In this context, icons have three­ dimensional shapes and textures. We reach out and touch three of the icons (academic discipline, cettainty, philosophic orientation), and off in a small corner we see them now labelling three orthogonal axes of a coordinate system labelled Knowledge, with tiny, illegible icons dotting that space. The Axes array has shrunk back down in size, and we now enter the Guides space. It grows, to become an array of pictures of people, arranged by their personality, profession, and teaching style. We select Jean, a blustery physicist who makes us draw our own conclusions. We notice that Jean is now standing off to one side of us, occasionally reaching over and peering at entries in the final, Filters array. It is filled with simulated clip-on sunglasses, with labels like NoAnalogies, EmphasizeTheory, and EmphasizeNames. These sunglasses filter out -- or stress -- various features of the tour we're about to take. We select NoMath and NoConnections, and put them on; Jean winces.

We tell Jean we're ready, and the new Knowledge array (with its axes labelled Field, Certainty, and Philos) · grows to fill our entire field of vision. In front of us, all around us, are all the entries in the knowledge base, organized by these three attributes. We notice that we're positioned ncar Physics on one axis (Jean's specialty), at NearCertainty on another axis (Jean knows I wouldn't have selected her to guide me around speculative fiction), and at Ontology along the third ax is (Jean prefers Teleology, or even Phylogeny, but knows I'm more interested in the current state of things). Moving around in this space for a couple minutes, with a running commentary from Jean, the image of Jupiter catches our eye, and we ask to spend some time finding out more about it

The Jovian sphere opens up, and inside is what appears to be miniature amusement park. It grows larger, until it appears almost life-sized and fills our field of vision. There's an arcade with pictures of the planet, a small library (presumably of books and arttcles on Jupiter), a scale model of the solar system in motion (with Jupiter highlighted), and a spaceship. Jean glances wistfully· at the scale model; if we hadn't put on the NoMath filter, she'd have shown us a simulation that demonstrates how the areas swept out by Jupiter in equal time periods are equal.

We browse through the textual material for a few minutes, occasionally pointing and asking Jean a question about what we've just read. Some of the books are labelled as original sources, but each one has a duplicate that's specially tailored and filtered for us, using a model of us, the clip-ons we're wearing, and even some more subtle and dynamic cues, such as our fidgeting, what specifically we've just read or skipped, etc. Since we chose to be on the Ontological plane, there is scarcely any material about what Jupiter was thought to be by the ancients, how it formed during the creation of the solar system, etc. Jean notices we're getting bored, and suggests we go in for a first-hand look. We board the rocketship, and after a visually and aurally believable liftoff find ourselves moving quickly into Jovian space. Jean knows that we prefer to control the ship ourselves, so conveniently sits in the co-pilot's seat.

One of our instruments, labelled Source, tells us that the images being generated out our viewscreen are taken from early NASA flyby mission photos. Jean answers our questions, and occasionally directs us to some interesting spots to fly over. When we insist on going down into the atmosphere, Jean scowls at a dial on the instrument panel, whose needle is dropping fast. Curious (we see the altimeter is elsewhere), we lean over and notice that dial is labelled Certainty, and the needle is moving from Fact to Fiction. Various other instruments report the chemical composition, pressure, temperature, etc. of the gases we are flying through. Suddenly, large whale-like creatures now begin to hurtle past us, and we pull a lever which returns us instantly to the Field/Certainty/Philos space. We note that we are indeed at the same Field/Philos point, but our Ceitainty has slipped to an extreme, namely fiction. In particular, we have emeq?;ed from a book-shaped icon labelled 2001/2010. Our anginal point of entry -- indeed, the entire encyclopedia of relatively certain knowledge -- is now far overhead along the Certainty axis. Pushing what appears to be a button on our watch, the session ends.

IT. Architecture of tl1c system

It is tempting to represent the architecture of Knoesphere as a single box containing the letters KB, for knowledge base, as even the inference procedures are represented within the same formalism and hence are a real, inspectable, modifiable part of the KB. Thus, this section is really about the architecture of the KB: the sorts of knowledge it will contain, how those are organized and used.

Of course a large part of the KB represents the encyclopedic knowledge: famous battles, buildings, biographies, birds, books, etc. The original text will be completely re-represented in a structured language, from which the program can generate text, or animation, or narration, or tactile sensations. But the decision as to modality is not made in advance. Regardless of output modality, the material being presented might range from a straight-forward re-representation of entries in the KB, to a dynamically-plotted story based on such data.

The knowledge base contains frames that represent heuristic rules. These deal with building and checking user models, appropriateness of various modalities and representations (including rules for when and how to cache, or pipeline, or expectation-filter). Some heuristics are domam-independent (learning strategies), but many are domain-specific; some heuristics generate plausible suggestions, some analyze and evaluate and critique [5,6].

Something which is absent from a typical encyclopedia but must be present in the Knoesphere KB is commonsense knowledge. This includes everyday physics [?,8], models of human interactions (including conversallon) [9], models of human/machine dialogues, as well ~ facts and heuristics about teaching, qucstion­ answenng, Imagery, analogy, etc.

. Much of that knowledge is usable tor an intelligent mterface to users, though additional details about people [10] are required: types of people (models of mathematicians, mailmen, and mechanical engineers) and individual people (celebrities, specific users, archtypical group representatives). Each such model includes physical appearance, goals, state of knowledge (including notatiOI_lS, vocabulary), memory capabilities (information proce~s.t~g 1_11odel of hu~an cognition), infercncing cap_abiitties, mterests an_d btases, how well they learn via vanous soits of teachmg methods, and miscellaneous attr~butes (strength, intelligence, etc.) Note that for each attnbute, we must record several values: the true value, the value the individual believes the true value to be the value the individual believes s/he is projecting to otl1ers, the valu_e tl1ey are projecting to others in truth, what they would hke the true value to be, what they would like to project to others, how strongly they want those last two values to become true (or to stay true).

Give'! that. the program has access to a very rich con~t:xt 111 w~tch the . user expressed something (eye positton, previous actions, model of what s/he is Interested in and typically asks about, where his/her finger is pointing), the tasks of understanding SP.eech and langua~e are constrained to the point of feasibility. The final ptece of the knowledge base deals with rules and algorithms for performing these tasks.

As implied above, the control structure is one of condition-action rules, executed in a forward direction, governed by a smaller and more general set of rules (actually a subset of the first set). These rules are represented in the same language as the rest of the knowledge, and are pmt of the KB, hence can be viewed and tailored (perhaps automatically [6]) to the individual user.

The user selects axes, a guide, and a set of filters. All ~ree choices serve the same function: cutting down the stze of _the space of information nodes to display, determmmg what structure to giv.e them in the display ?Jl:d. deciding where in that space to stait the user off mtttally. Often, the same constraint is available to a user in any one of the three choices he makes (e.g., s/he could enter a space where s/he chooses to be at Phylogeny ~ong one axis; or s/he could choose a guide who always mcludes !]le origins of the things he's presenting; or s/he could cltp on a filter called EmphasizeOrigins). A common preference may be for three axes that deal with subject matter, relegating all other constraints to the guide's personality and to the clip-on filters selected. Filters can be composed. For instance, one could take the NoConnections filter, and break knowledge up into separate concepts; putting the EmphasizeNames filter over that, followed by the EmphasizeOrdering filter, would result ~n an alp~abetized list of concept names, very much hke the mdex of a conventional print encyclopedia

There will be some handcrafted tours -- authored, just as ~Y work of art -- on file and accessible to the guide; this IS necessary because sometimes a sequence of exhibits is effective without anyone being able to account for why it's effective; si nce it can't be rcderived from more elementary planning principles, it must be preserved.

The user's responses are checked against the ex isting model of that ind ividual, and discrepancies are noted (and will eventually be reflected by changes in the model). This sort o f learning applies to models of groups of people as well ; thus, if mathematicians become less averse to examples over a period of years, the Mathematician model would slowly change to refl ect that.

Ill. Feasibility, Current Status, Future AI Issues

The project described in this paper is merely the expert systems "view" of an even larger effort, the design qf. ~. large "el.ectronic community" of services and facilities, that IS a focus of Atari's new Sunnyvale Research Center. Ignoring this larger vision, ignoring the unsolved problems with hardware and graphics and simulation-design, how feasible is the goal to represent [l]'s knowledge in an expert system?

Our efforts to date suggest that there will have to be one new. fr~me adde~ for each of t~e 300,000 paragraphs of matenal 111 [l]. Smce a very snml ar frame will usually ex ist, it wi ll take a person familiar with Knoesphere's data entry process (and having, say, a B.A. in the same field as the anicle) about one .hour to exploit the analogy and enter the new frame; this works out to three years of work for 50 such people. Analysis of [1]'s Macropaedia reveals only 400 distinct types of articles that we, the authors of this article, must represent by our target date of 1990. Unfortunately, there are approximately 2000 everyday concepts for us to represent (double the number in Basic English). We intend to spend much of the coming decade in research trying to build such a core.

In short, this "back of the envelope" calculation shows the representation pait of the project to be feasible (a decade of research, five years (and 200 man-years) of development). We will not be working on the problems of special hardware such as. helmet and gloves; high quality dynamic computer animation; home computers as powerful as current Lisp machines; inexpensive yet su:tably fast data links to central knowledge repositories; but several others at Atari and elsewhere are pursuing such technologies. The calculation, and the existing project to date, span only the task of re-representing the knowledge in [1], translating it from text into a frame­ based format that can be processed using existing expert systems techniques.

As mentioned previously, this project is in the initial stages, and most of the activity so far has been planning and destgn. We have taken some protocols of searching print ~nd current online encyclopedias to investigate strategtes people employ, and have constructed some sample simulations of the sort that will be included in the encyclopedia (though none as grandiose as that presented in . Sectiqn 1)._ The i~itial prototype ~ystem is being wntten m Ltsp, runnmg on Symbohcs LM-2 Lisp machines (soon to be 3600's) and Xerox D machines (1100s), using pieces of representation code from RLL [ll} and Eurisko [5,6], and simulation facilities from ThmgLab [12].

From our viewpoint as AI researchers, the project offers a unique testbed for context-constrained underst~nding. of speech, natural language, and other modahties which have not been programmable readily before (such as fa~ial .expressions and hand movements). It IS a world which IS almost as deep as most expe1t systems, and which IS 500 times broader. Analogy and metaphor play a key role . in human cognition, especially learnmg, but It has been dtfficult to experiment with those phenomena due to the narrow testbecls we have had availa ble previously. We will exploit analogy as to facilitate data entry, and to explain exhibits to the user; olher uses may be uncovered.

There is one final, crucial point to make. Tn the Abstract, we mentioned the Knocsphere's eventual impact on education, long after its introduction as a form of home entertainment. This delay is necessary because browsing through a "science museum" is not an effective way to learn science . . As enjoyable as it might be, the system presents facts, It doesn't teach principles. After it e~ists .. work can begin to add (probablx .manually) the effecttve teachmg programmes, the cntical analogies. Eventually, we hope that simulation-building tools will be good enough that, say, a physicist can expect to enter a simulation . su.ch as the Jovian example. This might well ~e done withm the framework of guides and filters, but it IS totally absent from current encyclopedias, and it will be largely absent from our initial system.

ACKNOWLEDGEMENTS

We thank Alan Kay -- and Atari -- for setting up an environment that provides the resources and freedom to undertake a project of this size. We also thank Charles van Doren and Bob Stein.