Comments on: Top-Down and Bottom-Up Semantics

By: Vipin Jain

Vipin Jain — Sat, 12 Jul 2008 21:04:52 +0000

No doubt. I still have to see a commercial AI, NLP, semantic system that works across seemingly infinite number of contexts and semantics from both users’ and information perspective. Leave alone that is not subjective .

Anyways, it was a fun discussion. And it was good to chat!

By: D Ashcart

D Ashcart — Fri, 11 Jul 2008 18:08:51 +0000

@Vipin – what you are saying makes sense. If you dig deeply enough into the origins of any learning system, you will hit the subjectivity barrier. Which is OK, since all forms of expression are ultimately subjective. This is the reason that the Holy Grail you describe is unreachable – you will need an infinite bitspace. You know you are headed into semantic swampland when you start seeing terms like ‘learning’, ‘domain expertise’, ‘heuristic’, ‘practical solution’, ‘intelligent’, etc. creep into the conversation. But there’s no denying that these solutions often offer utility.

My initial reaction to the article was to question the implied leap of faith from the untrustworthiness of a tagger to the trustworthiness of a corporation.

Nothing specific against Cognition – the semantic/NLP space is rife with marketing-speak-absolutism that tramples over nuance and accuracy, which occasionally tweaks my purist heart. This is my personal curse.

By: Vipin Jain

Vipin Jain — Fri, 11 Jul 2008 16:29:50 +0000

We have seen researchers positioned along the entire spectrum of AI and semantic web. Holy grail being a completely automated bottoms-up approach that requires no structured feed to the learning system – the system is supposed to learn the underlying structure on its own. This becomes extremely challenging when the learning system doesn’t have an identity, intelligence, or motivation of its own. Actually, it is extremely challenging when the learning system does have an identify, intelligence or motivation – just look at how good we are at training and retraining our children or other peoples .

Anyway back to computers, all learning systems have to be guided and differ only in the amount and means of guidance provided by the “algorithms specialist” or the “programmer”. Practical and successful approaches constrain the learning system and structure the problem such that simple mathematical rules can learn the classification boundaries from available data. Even then you can’t satisfy every user request in a given vertical. You optimize for the most common tasks that make a solution commercially viable from cost and ROI perspective. Our researchers at Retrevo have employed practical approaches of constraining problems by using intelligent crawling, statistics-based feature extraction and selection, and Bayesian learning of classification boundaries. We don’t intend to satisfy every corner case in our vertical equally well but we satisfy common tasks extremely well. Just 2 cents from a company that has applied research to solve practical problems with commercial viability. And we know we still have ways to go!

By: D Ashcart

D Ashcart — Thu, 10 Jul 2008 23:57:19 +0000

“An alternative method to “bottom-up” and “top-down” semantics is Semantic Natural Language Processing, such as that employed by Cognition Technologies, in which the computer has been taught the meanings and relationships of all of the words and phrases, and also recovers the meanings of the words and phrases in searched document set.”

I’m curious how this is an alternative to top-down. If top-down tagging is susceptible to the subjectivity and knowledge of the tagger, isn’t SNLP susceptible to the subjectivity of your programmers or librarians or ontologists?

In other words, who “taught the computer”?

By: Hope Leman

Hope Leman — Wed, 09 Jul 2008 13:29:54 +0000

This is a really fascinating article. I was interested that the emphasis was on coders as taggers. But isn’t a huge amount of tagging done by average people with no coding now-how at all? For instance, I help run a site that requires me to determine what categories each grant goes into it. Those are sort of quasi-tags that don’t require any coding know-how on my part and some of the grant listings do seem to get into Google (not as much as I would like, of course!).

And Blogger has the feature, “Labels for this post,” which is a sort of tagging system (I guess!). And then there is the whole world of del.icio.us and social bookmarking. None of that requires knowledge of tagging languages. And do websites necessarily need to tag their content if enthusiastic laypeople are doing it for them a la Digg?

Anyway, this was one of the best, clearest overviews of the very abstruse topic of the semantic web I have seen. Thank you. I am in library school and need to grasp all this stuff.

By: Andreas Harth

Andreas Harth — Wed, 09 Jul 2008 11:32:27 +0000

I don’t completely agree with the notion of top-down vs. bottom-up semantics. Even tagging individual pages with topics requires a formal specification of the topics, even if the formal spec is very basic, e.g. in the case of taxonomies such as DMOZ. The article also fails to explain that a shared use of URIs forms the basis for deriving meaning on the Semantic Web.

I’m skeptic of the promise that NLP methods will automatically derive meaning from text. While information extraction with NLP may work in narrowly specified domains (which, however involves a great deal of manual labour), automatically extracting entities, attributes and relationships from the open-domain, multi-lingual Web using NLP techniques has failed, despite millions of dollars of investment (Powerset). Rather, starting with structured datasets and extending and linking them iteratively seems to work (DBpedia, Linked Data).