Re: LDA Topic Model and Topic Maps

Aki Kivela <[email protected]> Thu, 04 Aug 2011 10:56:26 +0300
Newsgroups gmane.text.xml.xtm.general
Message-ID <[email protected]>
On 3.8.2011 22:41, Antonisamy Arokiasamy wrote:

> We have a LDA model that classifies documents and I am trying to see how
> this will map to Topic Maps and Ontologies.


Hi Antony

I am not really familiar with the LDA model but see the problem similar 
to any problem where a collection of documents is tagged with a set of 
keywords. Please correct, if I made a false interpretation. If right, a 
very simple approach is to

* Create a Topic Maps topic for each document
* Create a Topic Maps topic for each LDA topic
* Associate a document topic with a topic topic (if they should be 
interlinked, of course).

Now, if you have some additional information attached to the relation 
between the document and the topic, say a probability value, you can 
create third TM topic for each probability value and attach probability 
topic to the binary assiciation created above. As topic map associations 
may carry as many members as you like, you may add several additional 
"properties" to the association.

Alternative approach -- which is better if you want to address relations 
between LDA topics and documents directly -- is to

* Create a Topic Maps topic for each document
* Create a Topic Maps topic for each LDA topic
* Create a topic for each relation between the LDA topic and a document.
* Create an association between the relation topic and a topic topic.
* Create an association between the relation topic and a document topic.

This approach is more RDFish as you need only binary associations 
(triplets in RDF world) and probably better if have a lot of additional 
information you want to attach to the relation. The access and 
management of relations is more explicit but also takes more time and 
effort -- in my opinion.

I wonder if someone can figure out yet another modelling approach?

What I didn't explain above, is the details of transforming a document 
or a LDA topic into a TM topic. Neither didn't I touch the issue of 
choosing suitable vocabulary for required associations (i.e. what 
association type and what roles). As the topic map world is not that 
strict with vocabularies -- again, my opinion -- I suggest you invent a 
new vocabulary yourself unless you easily find a suitable one.

I hope I catched at least some useful points you can build upon.

Kind Regards,
Aki
Wandora Team


> Thanks
> /antony
>
>
>
>
> _______________________________________________
> topicmapmail mailing list
> topicmapmail-Zo64W7twoUFWk0Htik3J/[email protected]
> http://www.infoloom.com/mailman/listinfo/topicmapmail