On Sunday, Oct 12, 2003, at 16:13 Europe/Rome, Alan Gutierrez wrote:
The trouble with Wiki and docs is that new users, such as myself, are going to look for a documentation outline. A good TOC and index makes all the differnece in the world when searching documentation.
eheh, right on.
Has anyone discussed how to impose an outline on a Wiki?
yes. there are some proposals on the table, ranging from simple to futuristic.
- o -
the simple one is a manually created single-dereferencing linkmap.
Imagine that you have a repository with the following learning objects:
/1 /2 /3 /4
which are edited and created individually. Then you have a linkmap that basically says
Trail "A"
/1
Section "Whatever"
/3
/4
Section "Somethign else"
/2Trail "B" /4 /1
Trails are like "books", and they might share LOs. Trails might be published as a single PDF file for easier offline review.
Trails can be used as "tabs" in the forrest view, while the rest is the navbar on the side.
the LO identifier (http://cocoon.apache.org/LO/4) can be translated to a real locator (http://cocoon.apache.org/cocoon/2.1/A/introduction) and all the links rewritten accordingly.
This link translation is a mechanical lookup, based on the linkmap information.
[note: this is getting closer to what topic maps are about! see topicmaps.org, even if, IMO, it wouldn't make sense to use topic maps for this because they are much more complex]
- o -
The approach above works but it requires two operations:
1) creation of the LO 2) connection of the LO in the linkmap
It is entirely possible to have a page that lists all the LO, but it would be pretty useless. It is also possible to find a LO by searching... this is normally how wikis are accessed, but it would be better to find a way to have LO "self connecting".
- o -
In the more futuristic approach, each learning object specifies two things:
1) what it expects the reader to know 2) what it expect to give to the reader
the above reflects a mechanical model of the theory of cognition that indicates that information is transformed into knowledge by projection onto the existing cognitive context.
Suppose that you have learning object library of
/1 -(teaches)-> [a]
/2 -(teaches)-> [a]
[a] <-(expects)- /3 -(teaches)-> [b]
[a] <-(expects)- /4 -(teaches)-> [c]now you add a new learning object indicating
[c] <-(expects)- /5 -(teaches)-> [b]
and you can infer that
[a] <-(expects)- /4+/5 -(teaches)-> [b]
therefore
/3 and /4+/5
are cognitively related
- o -
The above is nice and good, but we hit the big wall: who defines the taxonomy and how.
The taxonomy is the collection of all those identifiers that [a] identify abstract concepts.
If each editor comes up with his own identifiers, we might have to issues:
1) identifiers are so precise that there is no match in between documents written by different people
2) identifiers are so broad in spectrum that concepts overlap and dependencies blur
This is, IMHO, the biggest problem that the semantic web has to face. The RDF/RDFSchema/OWL stack is very nice. A piece of art once you get it (and I still don't, but the fog is starting to clear up a little), but it's all based on taxonomical or ontological contracts... which are, IMO, the most painful and expensive thing to create an maintain.
So we must find a way to come up with a system that helps us in the discovery, creation and correctness maintenance of a taxonomy, but without removing or overlapping human judgment.
- o -
Here is a scenario of use (let's focus on text so, LO = page)
a) you edit a page.
b) you highlight the words in the page that you believe are important to identify the knowledge distilled in the page that is transmitted to the reader.
c) you also highlight the words in the page that you believe identify the cognitive context that is required by the user to understand this page
[this can be done very easily in the WYSIWIG editor, using different colors]
c) when you are done, you submit the page into the system.
d) the page gets processed against the existing repository.
e) the system suggests you the topic that might be closer to your page, and you select the ones you think they fit the best. or you introduce a new one in case nothing fits.
point d) seems critical in the overall smartness of the system, but I don't think it is since semantic estimation is done by humans on point e), point d) just has to be smart enough to remove all the pages that have nothing at all to do with the current learning object.
Possible implementations are:
- document vector space euclidean distance (used by most search engines, including google and lucene)
- latent semantic distance (but this one is patented but it will expire in a few years, used for spam filtering by the Mail.app in MacOSX, used by the Microsoft Office help system and the assistant).
- o -
The above model paints a double-dereference hypertext, sort of "polymorphic" hypertext where links are made to "abstract concepts" that are dereferenced against resource implicitly.
This allows the structure of the system and the navigation between LO to grow independently, using information from the various contextual dependencies and from the analysis of the use of the learning objects from the uses.
In fact, navigation between learning objects can be:
1) hand written 2) inferred from the contextual dependencies 3) inferred from the usage patterns
I believe that a system that is able to implement all of the above would prove to be a new kind of hypertext, much closer to the original Xanadu vision that Ted Nelson outlined in the '60s when he coined the term hypertext.
I don't think we have to jump all the way up here, there are smaller steps that we can do to improve what we have, but this is where I want to go.
-- Stefano.
