Difference between revisions of "Vulcan/PrototypeToWorking"

Latest revision as of 05:47, 13 September 2013

The following is a todo list for converting the prototype system into a working system that can operate on arbitrary input propositions. [Last update: Sep 12, 2013]

There are four main items:

• Switch to using real evidence instead of hand-written axioms. This means using the tuple database that Greg has built.
• Use external procedure calls to allow for dynamic verification of predicates. [Done]
• Writing a translator that can convert rules into Tuffy's MLN format.
• Integrate with textual evidence finder.

ETC: Friday, Sept 13th

Actual evidence

Open IE tuples

We need to handle nested relations and arguments with complex internal structure. (Examples here.)

• Open IE 4.0 handles n-ary but not nesting. [Michael will add post-processing to output nested tuples]
• We need a higher coverage version of Open IE that correctly handles internal structure in a complex argument.

Convert Open IE tuples into Tuffy's MLN axiom format.

Decide on stemming, head word extraction and other normalization to apply.

Import into postgres db.

WordNet

Convert WordNet RDF triples into Tuffy's MLN format.

Import into postgres db.

Compound Noun Categorizer (CNC) [Stopgap arrangement until Postgres external procedure is figured out.]

Process arguments (phrases) in propositions and study guide through CNC.

Extract relations that denote containment (e.g., "composed of")

Import into postgres db.

Figure out external procedures.[Done.]

Will do this early next week.

Rules

• Examples provide a starting point.
• Need to write a converter that translates human readable rules into MLN format.
• How to assign weights to the rules?

Propositions

• Extract best proposition.

Textual Evidence Finder

• Take output from TEF and convert it into a single MLN inference rule.