The economical (and successful) model for handling data of this sort is batch processing, possibly with some tool like Hadoop -- and our market research shows that many potential customers are already using this.
If we provided a live database to customers we'd have to make at least 10 times the revenue in order to cover variable costs and I'm not sure we could provide a service with satisfying performance at that level.
RDF technology is rapidly advancing and there's no substitute right now for a customer providing his triple store on his own machine with a lot of RAM and a few SSD drives.
The difference between SPARQL 1.1 and MQL (the propreitary Freebase language) is like the difference between chess and checkers.
In SPARQL you can write queries that involve any graph relationships that come into your head. You can take the UNION of multiple graph patterns. With SPARQL you can also get back 200,000 or more results.
With MQL you quickly run into the wall when you find that you can't write the query you want (you need to write 10,000 queries instead of one) or that the query you want to run times out, or that it only lets you get back a limited number of answers.
It's also possible to do batch processing with :BaseKB Pro with extreme efficiency. We did a calculation that would have taken 100 years on the biggest machine in the AMZN cloud in SPARQL in just 24 hours on a mac book pro with an optimized pipeline.
You'll need to process that file into something that your DB can consume. If all you need are minor syntax tweaks, you can stream straight through.
More likely, you'll need to rebuild the dump to be more than quads ordered by arbitrarily assigned unique id, and with nearly 600 million quads, it's not completely trivial to do, might involve temporary index files, etc.
It's a machine readable encyclopedia with very good coverage of concepts that are in people's shared conciousness.
If it were combined with a large document corpus (say CommonCrawl) you could use it to build (i) something like DMOZ that needs little human input, or (ii) a document analyzer that extracts named entities and topics, (iii) a knowledge base to support "enterprise search" where the major problem is recall because people don't use the right keywords, or (iv) many other intelligent apps which I haven't thought of yet.
Freebase isn't the same thing as Google Knowledge graph. Freebase contains 22M entities (according to the wikipedia entry), while the knowledge graph contains 500M entities (according to google announcement).
Actually the 500M (or 400M or 300M) or whatever number is bogus.
We do know that they ported "graphd", the Freebase database, to Google's infrastructure. Really nothing happened at Freebase for a year and a half while they were doing this.
They've probably created a very large Freebase graph for scale testing and it's likely that they're able to handle 500M entities from a hardware perspective.
If you look at how often the GKG comes up in search, it clearly doesn't come up very often. So the process that makes the visible GKG is a process of subtraction, not addition. There are many topics where the GKG could give you an answer from Freebase but it doesn't, and they prune many of them away because they don't want to take the chance of giving wild answers. It's more likely that 1 or 1.5 million topics are in the visible GKG.
I'm not at liberty to go into much detail here, but there are a number of things that are in the KG that aren't in freebase. The process giveth and the process taketh away.
> Actually the 500M (or 400M or 300M) or whatever number is bogus.
Could you provide some sources here?
> They've probably created a very large Freebase graph for scale testing and it's likely that they're able to handle 500M entities from a hardware perspective.
If this were reality it would be a pretty big lie on Google's part. Why do you think this is the case?
EDIT: OK, now I realize why your comments trigger my BS alarm. It's in your interest to represent the Google knowledge-graph as being equivalent to Freebase/metaweb (a Google acquisition many years ago. Do you really think they did nothing to increase their graph since that acquisition?) because you are the founder of Ontology2 (the company providing this service). You really should mention that somewhere in this thread or on your profile (I realize your comments elsewhere in this thread allude to it -- for the claims you are making without sources that's not enough IMHO).
The Knowledge Graph is growing quite quickly as we add more and more data sources. Freebase is one of those sources but there are obviously others that make up the full 500M entities.
The rate at which we add entities to the Knowledge Graph is different from the rate at which those entities appear in search. What you see in search now is just part of the Knowledge Graph but we're continuously increasing the coverage while making sure that the quality of the data remains extremely high. I guess that's what you would call the "visible GKG"; to us, its all one big graph.
(i) it would only let you look at one subject at a time,
(ii) it would only return a maximum of 100 facts about a subject
(iii) it didn't use names consistently so you couldn't match up properties and classes with the schema
It's all great for a hack day demo but not acceptable at all if you care about building systems that give the right answer.
Great stuff! On the landing page, I would add some examples of what one can do with this KB.
e.g. a startup in local space can use BaseKB to find various attributes about a city and nearby towns etc. You can also price various subsets geared to such niches.
Most startups re-invent the wheel when it comes to such KB/data issues. Show them the math :- instead of spending 1/0.5 FTE time in aggregating/cleaning data, they might as well shell out $199pm and get the job done.
DBpedia is produced by a rube goldberg machine that parses (undocumentable) wiki markup
if a fact is wrong in Dbpedia you need to do detective work to know if it's a problem with the mappings or with Wikipedia. If it's in Wikipedia maybe the editors will let you fix it on a good day.
Precision and recall for types like "Person" is much better in Freebase. Years ago I spent a week in a half cleaning up data from Dbpedia to do something I was able to do in 15 minutes with Freebase.
It should be noted that the RDF endpoint isn't the only way of getting data out of Freebase.
Freebase provides complete weekly data dumps (http://wiki.freebase.com/wiki/Data_dumps) and has instructions for post-processing them; presumably this is exactly what PaulHoule is doing. As chintan pointed out, it could still be a worthwhile service for you.
26 comments
[ 4.3 ms ] story [ 62.6 ms ] threadIf we provided a live database to customers we'd have to make at least 10 times the revenue in order to cover variable costs and I'm not sure we could provide a service with satisfying performance at that level.
RDF technology is rapidly advancing and there's no substitute right now for a customer providing his triple store on his own machine with a lot of RAM and a few SSD drives.
In SPARQL you can write queries that involve any graph relationships that come into your head. You can take the UNION of multiple graph patterns. With SPARQL you can also get back 200,000 or more results.
With MQL you quickly run into the wall when you find that you can't write the query you want (you need to write 10,000 queries instead of one) or that the query you want to run times out, or that it only lets you get back a limited number of answers.
It's also possible to do batch processing with :BaseKB Pro with extreme efficiency. We did a calculation that would have taken 100 years on the biggest machine in the AMZN cloud in SPARQL in just 24 hours on a mac book pro with an optimized pipeline.
More likely, you'll need to rebuild the dump to be more than quads ordered by arbitrarily assigned unique id, and with nearly 600 million quads, it's not completely trivial to do, might involve temporary index files, etc.
Still, it's doable.
(also see: http://code.google.com/p/freebase-quad-rdfize/ )
SPARQL is harder to learn and set up, but allows for more complicated queries that might require offline processing.
(I designed the first version of MQL)
>just about anything that would come up in a game of "20
I don't feel like that is clear enough.
If it were combined with a large document corpus (say CommonCrawl) you could use it to build (i) something like DMOZ that needs little human input, or (ii) a document analyzer that extracts named entities and topics, (iii) a knowledge base to support "enterprise search" where the major problem is recall because people don't use the right keywords, or (iv) many other intelligent apps which I haven't thought of yet.
We do know that they ported "graphd", the Freebase database, to Google's infrastructure. Really nothing happened at Freebase for a year and a half while they were doing this.
They've probably created a very large Freebase graph for scale testing and it's likely that they're able to handle 500M entities from a hardware perspective.
If you look at how often the GKG comes up in search, it clearly doesn't come up very often. So the process that makes the visible GKG is a process of subtraction, not addition. There are many topics where the GKG could give you an answer from Freebase but it doesn't, and they prune many of them away because they don't want to take the chance of giving wild answers. It's more likely that 1 or 1.5 million topics are in the visible GKG.
Could you provide some sources here?
> They've probably created a very large Freebase graph for scale testing and it's likely that they're able to handle 500M entities from a hardware perspective.
If this were reality it would be a pretty big lie on Google's part. Why do you think this is the case?
EDIT: OK, now I realize why your comments trigger my BS alarm. It's in your interest to represent the Google knowledge-graph as being equivalent to Freebase/metaweb (a Google acquisition many years ago. Do you really think they did nothing to increase their graph since that acquisition?) because you are the founder of Ontology2 (the company providing this service). You really should mention that somewhere in this thread or on your profile (I realize your comments elsewhere in this thread allude to it -- for the claims you are making without sources that's not enough IMHO).
The rate at which we add entities to the Knowledge Graph is different from the rate at which those entities appear in search. What you see in search now is just part of the Knowledge Graph but we're continuously increasing the coverage while making sure that the quality of the data remains extremely high. I guess that's what you would call the "visible GKG"; to us, its all one big graph.
I go around saying there's no ring to bind them all but that an arbitrarily good aproximation can be built.
That said, this looks very interesting and I expect businesses to grow by selling public data in easier to use forms.
(i) it would only let you look at one subject at a time, (ii) it would only return a maximum of 100 facts about a subject (iii) it didn't use names consistently so you couldn't match up properties and classes with the schema
It's all great for a hack day demo but not acceptable at all if you care about building systems that give the right answer.
e.g. a startup in local space can use BaseKB to find various attributes about a city and nearby towns etc. You can also price various subsets geared to such niches.
Most startups re-invent the wheel when it comes to such KB/data issues. Show them the math :- instead of spending 1/0.5 FTE time in aggregating/cleaning data, they might as well shell out $199pm and get the job done.
DBpedia is produced by a rube goldberg machine that parses (undocumentable) wiki markup
if a fact is wrong in Dbpedia you need to do detective work to know if it's a problem with the mappings or with Wikipedia. If it's in Wikipedia maybe the editors will let you fix it on a good day.
Precision and recall for types like "Person" is much better in Freebase. Years ago I spent a week in a half cleaning up data from Dbpedia to do something I was able to do in 15 minutes with Freebase.
Freebase provides complete weekly data dumps (http://wiki.freebase.com/wiki/Data_dumps) and has instructions for post-processing them; presumably this is exactly what PaulHoule is doing. As chintan pointed out, it could still be a worthwhile service for you.