Logical Inferencing with normal relations in DeepDive

[Kaushik Chakraborty]

Hi,

How to do pure logical inferencing in DeepDive on normal relations. Documentations and examples only show probabilistic inferencing with first-order-logic operators to co-relate variable relations. In a Knowledge Base creation logical inferencing is important to derive new knowledge from the facts on the basis of logical rules. Having a DataLog like syntax gave us a premonition that we can write prolog-like rules to encode logical relations like  transitivity & reflexivity. Normal derivation syntax did not help realise the same as it's not processing the rules recursively or right away rejecting in the case when relations appear both in the head and body. Trying to use logical operators like implication & conjunction with normal relations also gave error. 

Is there some way to realise these aspects with DeepDive's syntax or the only ways are either try to do probabilistic inferencing or write an UDF ?

[Jaeho Shin]

Hi Kaushik,

Leaving recursion out of DDlog was a design choice we made for simplicity.  Although it's a bit ugly you can encode recursion up to a finite limit. (See: https://gitter.im/HazyResearch/deepdive?at=5822a923c2f2cf7275e9e712)

The "inference rules" in DeepDive/DDlog is only for probabilistic inference.  All logical/deterministic rules have to be applied ahead of time.  In my view, you can apply deterministic / logical inference rules by either encoding them as normal derivative rules in DDlog, or using other means (UDF or SQL) to materialize data implied by logical inference to make sure the rules are reflected in the final factor graph for probabilistic inference.

Hope this helps,

~Jaeho

--
You received this message because you are subscribed to the Google Groups "deepdive-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to deepdive-user...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/deepdive-users/a7d7ed87-3fde-4943-bf87-6e40ddd4a148%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.