Expert System Brings Semantic Smarts to Advertising
Expert System is a perhaps little known "semantic intelligence" company; but it has 15 years of experience in the tech industry, 115+ employees and is bringing in a very solid $12 Million a year in revenues from over 100 corporate and government clients (at 40% growth over the past two years). The Italian company's core technology is the Cogito platform, a sophisticated system which searches, extracts and classifies unstructured information and makes it into structured data. Cogito (which translates to "I think" in Latin) is bringing semantic technologies to the mainstream commercial world, including online advertising.
We spoke to Brooke Aker, CEO of the US subsidiary of Expert System, to find out more about the underlying technology of Cogito and its commercial applications. In particular we talked about how Expert System is using semantic technologies to power a new type of advertising.
The basic premise of Cogito is that it transforms unstructured information into structured data. Out of this process comes a "semantic network", which is much the same thing as what rival company Cognition called a "semantic map" in our September '08 interview with them. It's important to point out that Cogito isn't necessarily a 'Semantic Web' application, but it does use things like natural language processing and other semantic analysis. It can output RDF, but that isn't a fundamental part of Cogito.
Brooke Aker described the system to us as being like an "electronic dictionary". There are 350,000 words and 2.8 million "relationships" in Cogito. Cognition claimed 10 million "semantic connections" in its map back in September, but Aker suggested that it wasn't quite an apples and oranges comparison. According to Aker, Cogito's semantic network is "richer" than Cognition's.
Expert System's new semantic advertising solution, Cogito Semantic Advertiser (CSA), came about because the company saw that traditional contextual keyword advertising is resulting in a lot of inaccuracies and mistakes when matching ads to page content. The classic example is the jaguar one, where a story about a jaguar (the animal) is accompanied by a 'contextual' advert for jaguar the car. Expert System told us that this kind of scenario won't occur with their semantic advertising system.
Expert System claims to have come up with an advertising solution that understands "the real meaning" of words, based on theories of human comprehension. For example, their system analyzes the semantic meaning of words and their context. So in the jaguar example, Expert System would 'understand' that jaguar is an animal in the context of the story - and therefore it would not serve up ads about the jaguar car.
One feature of Cogito Semantic Advertiser that stood out for us was the granular categorization, which allows for very fine grained targeting of ads. Brooke Aker told ReadWriteWeb that their product has already created around 2 million niches for advertisers to target, which is something that many Long Tail publishers will find appealing.
We're impressed by the semantic software that Expert System is producing, not just with advertising but in other commercial and government applications. Let us know in the comments if you've come across this company before and if so, what your thoughts are.
Share
Comments
Subscribe to comments for this post OR Subscribe to comments for all ReadWriteWeb posts
Hi Richard,
Thanks so much for sharing. I believe strongly in the semantic web, and have Xerox Parc and HP Labs research backing our technology. Looks like this company provides another great alternative for our semantic targeting.
I've forwarded this article to my team.
Cheers,
Marissa
Google would better buy COGITO in order ot secure their advertising position.
These kind guys from Modena (yeah, just where Ferrari are made!) demoed their solution in my company a few months (>18) ago.
One word: impressive!
But as they honestly stated, the "semantic" gets "smart" if you train it! And training can require a lot of work.
The thousand or millions of links between word are not automatically generated, and must be defined on a "subject" basis, meaning that changing subject can require a small|quite large|very large rework.
So what I miss (I told: "I miss", me, not the product) is something that helps building these links.
Great product and great company anyway.
Eugenio says we need something to help build the links..I agree.
So what would that be exactly? An ontology sand box, a reasoner/inference engine and some kind of checking tool + RDF and triples and microformats perhaps some FOAF thrown in?
What I can't quite understand is how exactly we are supposed to throw these things all together to make a semantically well formed doc with machine interoperability.
If someone could set up a visual model and combine it with a way to connect with people across the web who are really keen on scaffolding semantic documentaria (see: www.innocentive.com) that included a public sand box like you see in Second Life perhaps we could get some semantically sparkling results and energize people to contribute from across web communities.
What I mean is: we could get contributions from non techies and from techies. People would become inspired to work on semantic projects outside medicine + bio.
I think OWL tutorials are beyond the beyond of boring.
And it would help if we could see a glimmer of how semantic mark up can lead to time savings and improved customer service for web marketing types.
Best,
HAL
Actually Cogito knowledge base is already "trained" for general purpose and most applications. Eugenio is absolutely correct when he says that some training is needed, but only when you're about to target specific matters (for instance medicine, petrochemistry, etc.).
Phil, sure. You, ehm, the system needs more training related to the specific matter. Or a specific language, right?
I need to tell Cogito "door" is different from "brings", even if in italian they are both translated with "porta" (I'm still looking for an English language example....).
I'm far from expert in the subject, but I suppose that part of the training can be done feeding the system with a lot of text. But I have no idea on what percentage of the training needs to be done/tailored with "manual" work.
HAL, being "far from expert" makes me able to answer: I don't know. :)
I'm just guessing, but I think the solution lies somehow in the "social" effect, i.e. users vote implicitly best results/matches. The only (technically) viable alternative (formal OWL wrapping of the subject) is, as you say, far from reality.
Excuse me if I misused some term. Again, this is not my subject.