On Tue, Feb 23, 2016 at 1:28 AM, Markus Krötzsch <markus@semantic-mediawiki.org> wrote:
On 22.02.2016 18:28, Tom Morris wrote:
On Sun, Feb 21, 2016 at 4:25 PM, Markus Krötzsch
 <markus@semantic-mediawiki.org <mailto:markus@semantic-mediawiki.org>>
wrote:
    On 21.02.2016 20 <tel:21.02.2016%2020>:37, Tom Morris wrote:
        On Sun, Feb 21, 2016 at 11:41 AM, Markus Krötzsch
        <markus@semantic-mediawiki.org
    
        <mailto:markus@semantic-mediawiki.org>>>
        wrote:

             On 18.02.2016 15:59, Lydia Pintscher wrote:

                 Thomas, Denny, Sebastian, Thomas, and I have published
        a paper
                 which was
                 accepted for the industry track at WWW 2016. It covers
        the migration
                 from Freebase to Wikidata. You can now read it here:
        http://research.google.com/pubs/archive/44818.pdf

             Is it possible that you have actually used the flawed
        statistics
             from the Wikidata main page regarding the size of the
        project? 14.5M
             items in Aug 2015 seems far too low a number. Our RDF
        exports from
             mid August already contained more than 18.4M items. It
        would be nice
             to get this fixed at some point. There are currently almost 20M
             items, and the main page still shows only 16.5M.

        Numbers are off throughout the paper.  They also quote 48M
        instead of
        58M topics for Freebase and mischaracterize some other key
        points. They
        key number is that 3.2 billion facts for 58 million topics has
        generated
        106,220 new statements for Wikidata. If my calculator had more
        decimal
        places, I could tell you what percentage that is.

    Obviously, any tool can only import statements for which we have
    items and properties at all, so the number of importable facts is
    much lower.

Obviously, but "much lower" from 3.2B is probably something like
50M-300M, not 0.1M.

That estimate might be a bit off. The paper contains a detailed discussion of this aspect.

Or the paper might be off. Addressing the flaws in the paper would require a full paper in its own right.

I don't mean to imply that numbers are the only thing that's important, because that's just one measure of how much value has been extracted from the Freebase data, the relative magnitudes of the numbers are startling.
 
The total number of statements that could be translated from Freebase to Wikidata is given as 17M, of which only 14M were new. So this seems to be the current upper bound of what you could import with PS or any other tool.

Upper bound using that particular methodology, only 4.5M of the 20M Wikidata topics were mapped when, given the fact that Wikidata items have to appear in a Wikipedia and that Freebase include all of English Wikipedia, one would expect a much higher percentage to be mappable.
 
The authors mention that this already includes more than 90% of the "reviewed" content of Freebase that refers to Wikidata items. The paper seems to suggest that these mapped+reviewed statements were already imported directly -- maybe Lydia could clarify if this was the case.

More clarity and information is always welcome, but since this is mentioned as a possible future work item in Section 7, I'm guessing it wasn't done yet. 

It seems that if you want to go to the dimensions that you refer to (50M/300M/3200M) you would need to map more Wikidata items to Freebase topics in some way. The paper gives several techniques that were used to obtain mappings that are already more than what we have stored in Wikidata now. So it is probably not the lack of mappings but the lack of items that is the limit here. Data can only be imported if we have a page at all ;-)

If it's true that only 25% of Wikidata items appear in Freebase, I'd be amazed (and I'd like to see an analysis of what makes up that other 75%).
 
Btw. where do the 100K imported statements come from that you mentioned here? I was also interested in that number but I could not find it in the paper.

The paper says in section 4, "At the time of writing (January, 2016), the tool has been used by more than a hundred users who performed about 90,000 approval or rejection actions." which probably means ~80,000 new statements (since ~10% get rejected). My 106K number is from the current dashboard.

Tom