NLP to RDF to VSM ?

[stcruy]

Moving an email conversation to here, in the public interest:

–> Idea:

I have looked at your overview image and read some of the linked publications. And I can see the identified steps.
Another point or path could be to translate the text dependency trees into RDF (including mapping to ontology and terminology, I could imagine that for this step some publication exists) and then translate to VSM-json? Make this sense? Complexity? Losing context?

–> Reply:

Indeed, clever, that may be a path worth exploring.

Some thoughts:

1. a tool that could link terms (words/phrases) to URIs would be a great help. (Enju does some already, I think).

2. for conversion of the structure: it depends on what such existing RDF-exporters' output would represent though:
   • If dependency relations (nsubj, aux, etc.) are only mapped literally to RDF predicates, then no 'semantic upgrade' would really happen yet.
   Still, it would immediately enable the use of RDF graph substitution tools in a next step, I'd think.
   • Btw, such literally converted RDF would be distinct from what the image's RDF stands for.
   The image's RDF is positioned after the semantic upgrade.
   It would correspond to a RDF-VSM schema (using only 'subject' and 'object' predicates, and a couple of others). (Formalizing such a schema is still a todo).
   That VSM-schema'd RDF would serve as a faithful storage format for the VSM. And it would directly support queries that fish for VSM-substructures.
   • If any tool could map dependency trees onto the "N-ary relations" RDF paradigm (and deeply so), that could be a great start!...

Btw I came across a newly started multi-year project 'Universal Natural Language Understanding' that aims map text onto a formal-logics form (Discourse Representation Theory) ...which is a rather complicated form though. They will try out both rule-based methods and machine-learning. (I sent them a message about VSM yesterday).
Point is: the hackathon may just be the start of an ambitious project :)