Comments on: Semantic vectors and bounded subjects

By: Chris

Chris — Sat, 13 Aug 2011 06:56:14 +0000

Interesting – and crossed a discussion had yesterday needing a similar solution – as a 2 year old thread it’s a bit of a shame you haven’t revisted the topic to indicate if you have made any progress in finding any pointers or solutions.

By: Beth

Beth — Tue, 23 Jun 2009 14:06:08 +0000

“With semantic web technologies, ontologies and knowledge bases – can it be so hard to use services like Twine, True Knowledge etc etc to build subject maps.”

True Knowledge is focussed on giving straight answers to direct questions, so the whole ontology is geared towards understanding exactly what you asked and giving you an exact answer.

Having said that, it would be possible to build subject maps out of the True Knowledge ontology if there was someone who wanted to use them.

By: Peter C

Peter C — Sat, 18 Apr 2009 02:21:31 +0000

You want an introduction to a topic? Kosmix is the daddy of this stuff.

Dunno what content you *expect* but Kosmix generated this: http://www.kosmix.com/topic/Coagulation_measurement_devices?

Kosmix utilizes semantic technologies to generate subject maps. It’s a one stop shop for aggregated knowledge. Like Mahalo, but not dumb (Mahalo is edited by humans)

By: Laurence

Laurence — Fri, 17 Apr 2009 10:48:59 +0000

very cool… lets chat for sure. Prof Zoubin had some “this is most similar to that” or “if this is the sub set… then this is the super-set” algorythms and so I wonder if this is adjacent to the solution set we need. It strikes me that Twine is a cluster of people who define knowledge spaces semantically… could this be used as a primer.
I’ll buy you a coffee..

By: Martin Kleppmann

Martin Kleppmann — Mon, 13 Apr 2009 08:18:09 +0000

I also find this an extremely interesting area, but unfortunately also a very difficult one as it goes deep into the heart of AI. It is even tricky to nail down what “a topic” is, and how it can have sub-topics, super-topics, contrasting topics, cross-domain linked topics, and topics which are related to some degree or other. I have discussed similar ideas with a number of people, and have written a research proposal for a slightly more manageable problem: analysing the relationships between people and topics. People are well-defined entities, making them easier to analyse; topics and subjects attach themselves naturally to people (interests) and their interactions with other people (conferences, citations, joint publications, discussions etc. on a particular topic). And if you can find out who the key opinion leaders are in a field, and also the rebellious outsiders, you can just go and talk to them directly (that’s the “it doesn’t matter how little I know, as long as I know someone who does know” approach).

A workable solution of this problem would probably try to extract facts from natural language text using computational linguistics and machine learning (e.g. recognising phrases like “in contrast to Y, X is A while Y is B and C). Anything learnt this way will of course be probabilistic/bayesian, not definite. This can be augmented with information from the semantic web, but I don’t think it’s possible yet to rely solely on manually-built semantic data sources such as ontologies as they are still quite limited.

I am planning to work on this under the umbrella of a PhD with the natural language processing people at the Cambridge Computer Lab, in collaboration with Zoubin in Engineering. I think this would be the right environment to learn the state-of-the-art techniques and to start building a prototype, which could then be spun out into a business when the time is right.

Happy to discuss!