Just an idea of a very sparse but hopefully not so bad encoding (I have not actually
tested it).
NB: I am going to use a lot the terms defined in the glossary [1].
A value could be encoded by a vector:
- for entity ids it is a vector V that have the dimension of the number of existing
entities such that V[q] = 1 if, and only if, it is the entity q and V[q] = 0 if not.
- for time : a vector with year, month, day, hours, minutes, seconds, is_precision_year,
is_precision_month, ..., is_gregorian, is_julian (or something similar)
- for geo coordinates latitude, longitude, is_earth, is_moon...
- string/language strings: an encoding depending on your use case
...
Example : To encode "Q2" you would have the vector {0,1,0....}
To encode the year 2000 you would have {2000,0..., is_precision_decade =
0,is_precision_year=1,is_precision_month=0,...,is_gregorian=true,...}
To encode a snak you build a big vector by concatenating the vector of the value if it is
P1, if it is P2... (you use the property datatype to pick a good vector shape) + you add
two cells per property to encode is_novalue, is_somevalue. To encode "P31: Q5"
you would have a vector V = {0,....,0,0,0,0,1,0,....} with 1 only for V[P31_offset +
Q5_offset]
To encode a claim you could concatenate the main snak vector + the qualifiers vectors that
is the merge of the snak vector for all qualifiers (i.e. you build the vector for all snak
and you sum them) such that the qualifier vectors encode all qualifiers at the same time.
it allows to check that a qualifiers is set just by picking the right cell in the vector.
But it will do bad things if there are two qualifiers with the same property and having a
datatype like time or geocoordinates. But I don't think it really a problem.
Example: to encode the claim with "P31: Q5" main snak and qualifiers "P42:
Q42, P42: Q44" we would have a vector V such that V[P31_offset + Q5_offset] = 1,
V[qualifiers_offset + P42_offset + Q42_offset] = 1 and V[qualifiers_offset + P42_offset +
Q44_offset] = 1 and 0 elsewhere.
I am not sure how to encode statements references (merge all of them and encode it just
like the qualifiers vector is maybe a first step but is bad if we have multiple
references). For the rank you just need 3 booleans is_preferred, is_normal and
is_deprecated.
Cheers,
Thomas
[1]
https://www.wikidata.org/wiki/Wikidata:Glossary
Le 27 sept. 2017 à 12:41, John Erling Blad
<jeblad(a)gmail.com> a écrit :
Is there anyone that has done any work on how to encode statements as features for neural
nets? I'm mostly interested in sparse encoders for online training of live networks.
_______________________________________________
Wikidata mailing list
Wikidata(a)lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wikidata