How can I use deepdive to perform unsupervised verification?

[rudaoshi]

Hi, I have a unlabeled data set in which each sample has an initial confidence. I want to do some verification by writing rules about commonsense to reduce the confidences of samples which violate commonsense. How can I do this?

1. How can I use the initial confidence?
2. How can I assign weights to the commonsense rules?

Thank you very much!

[rudaoshi]

When I want to define a unsupervised model by using "p(x,y)=NULL: ....", the program report:

column "label" is of type boolean but expression is of type text

It seems that the error is about a bug in sql generation module.

[chrismre]

You may want to check out http://hazyresearch.github.io/snorkel/, which is more directly about weak supervision. DeepDive can do all this (and much more!). This flexibility means that not all representation decisions are obvious. Chris

…

On Mon, Dec 19, 2016 at 8:24 PM 孙明明 ***@***.***> wrote: When I want to define a unsupervised model by using "p(x,y)=NULL: ....", the program report: column "label" is of type boolean but expression is of type text It seems that the error is about a bug in sql generation module. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#608 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHPtuEFo-aPa3pwU-pGppjuv12ulh3Pfks5rJ1iIgaJpZM4LRaM2> .

[rudaoshi]

@chrismre Thank you for mentioning another amazing projects. I'll look at it.

However, currently I'd rather like to know how to do this by deepdive.

[chrismre]

DeepDive can express essentially any factor graph. You'll need to write the rules that create the required factor graph. A default factor graph for this process is described in the Snorkel/data programming paper.

Hope that helps! Chris

[alldefector]

@rudaoshi Here is an example rule with manually assigned weights (adapted from the "census" example):

@weight(1.2)
rich(id) :- adult(id, _, workclass, _, _, _, _, _, _, _, _, _, _, _, _, income_bracket).

You can change the rule body (the adult... part) to encode the "common sense" cases.

[thodrek]

@rudaoshi Regarding the NULL issue alone, you should try to use casting within ddlog. Writing = NULL::boolean instead of = NULL will take care of the label casting issue.

[rudaoshi]

Thank you very much to every one. I'll try your recommended solutions.

[rudaoshi]

@thodrek I got following error using Deepdive 0.8:

2016-12-21 11:43:00.679659 [error] app.ddlog[176.44] failure: :-' expected but :' found
2016-12-21 11:43:00.679741
2016-12-21 11:43:00.679756 has_relation(p1_id, p2_id, relation) = NULL:boolean
2016-12-21 11:43:00.679767
2016-12-21 11:43:00.679947 ^

Does this feature need newer version cloned from github?

[alldefector]

@rudaoshi note that there are two colons: NULL::boolean. If that doesn't work, try cast(null as boolean).