Re: Inconsistent models, mapping, interoperability, and the SUO

["Rich Cooper" <richcooper@mindspring.com>]

Rob Freeman wrote:
> On Friday 25 March 2005 23:49, Rich Cooper wrote:
>> Rob Freeman wrote
>>
>> > Most of the information you need is not captured in those (assumed)
>> > hierarchical relationships though, is it?
>>
>> Correct, that information is not in WordNet 2.0 and also not all of the
>> drugs that have to be watched for are listed in WordNet either.
>>
>> The point is that WordNet contains the initial charge of common
>> English terms, a lattice of nouns, the synsets of nouns, verbs,
>> adjectives
>> and adverbs in common usage.  It also contains a number of phrases,
>> definition glosses, and other useful material.
>
> OK, but you are starting with something which does not have (all) the
> information you need, and which is going to force a (weak) format on the
> information you need when you do get it
>
> What I want to see is information in the structure. If we have information
> in
> the structure we can restructure it to suit our purpose. That this is
> necessary I hope has been established over the last few weeks
> (philosophical
> interpretation apart.)
>
> The fewer preconceptions there are in the actual nodes, and the more in
> the
> structure, the more powerful (capable of reinterpretation) your
> representation will be.
>
> WordNet gives you some structure, but as far as I can see it also takes it
> upon itself to represent a lot of information (most?) where it can't be
> reinterpreted, in the nodes.
>
> And above all from your point of view, it doesn't even have most of the
> information you need.
>
> Why not do a little bit of extra early work to put some machine learning
> infrastructure in place? Then you can extract all the information you need
> from texts. And as a bonus you can keep that information in the structure,
> which gives you the power to reinterpret.

Machine learning of English statements requires both true and false examples
of each concept.  That means a lot of labor in developing the training
samples.
In normal medical records texts, there shouldn't be any false examples,
except
by accident or error on the part of the transcriber.

So machine learning doesn't leverage human effort in this case as much as it
might in other areas.

But using WordNet class structures at least gives structure to
generalizations
and specializations of sentences based on replacing one member of a class
with some other member of the same class.  Knowing the class names and
the class members is very valuable in reducing human effort required for
this.


> I don't know how you would extract drug interaction information directly
> from
> texts, but I'm sure it's possible (because I believe meaning is coded in
> texts in patterns, which can be found.)
>
> Can you give me some examples of exactly the kind of information you need,
> in
> a structure you could use.
>
> -Rob

I can't get too specific due to proprietary nature of the application.  I
want to
be able to separate the tools from the application, and getting too specific
works against that goal.

Rich