Vulcan/MeetingNotes/Aug23 2013

Notes

• Implement the prototype system that would work on 3 questions. [ETC Aug 27, 2013]

• Assume the knowledge and axioms necessary.
• Extract as much useful debug information from the system as possible. (e.g., The set of axioms and rules used, new facts that were inferred.)

• Stephen will a) check if the patterns will work on the larger definitions corpus, and b) investigate prior work on definitions processing.

• Niranjan will send Stephen pointers to the data and Peter's summary of prior work.

• The propositions are nested structures e.g., (Growth, causes, (leaves of a plant, become bigger)). Open IE 4.0 does not handle nested tuples as of now. We need address this mismatch.

• Oren's suggestion is to collect examples of required representations from the data and compare them to existing output from Open IE 4.0 to figure out how to fix the mismatch.
• Have a meeting at a later time with Mausam and Michael.

• Greg will aim to build a first version of the Textual evidence finder by next Friday.

Update

• Created a plan for a prototype system that will work on three example questions.

• Stephen helped design a representation for the rules. We will have human-authored rules that will be translated to Tuffy's MLN format.
• See documents on rule representation and transforming to MLN syntax for details.

• Weak evidence finder.

• Greg is getting started on weak evidence finder.
• Started a specs page for this component.
• Greg will flesh out the details.

• How to perform procedural escapes within Tuffy? Tuffy doesn't support procedural escapes but we can make do with some hacks.

• Tuffy allows built-in postgres functions in inference rules.
• We can write our own postgres procedures (e.g., stemmer, head word extractor etc.) that can be used as a function in the inference rule.
• Tuffy developers think this should be doable easily.

• Other MLN Systems:

1. RockIt (Mathias Neipert and co.) -- Quite similar to Tuffy but is faster (supposedly). Seems to have more convenient notations for representing textual arguments. Not easy to have constraints on variables used in rules. Don't know if this system can support procedural escapes easily. Will find out.
2. theBeast (Sebastian Reiedel) -- Representations are suited for textual arguments but system does not support marginal inference (i.e., computing the probability of a query predicate).

• Get a knowledge spec:

• isa, partOf, etc.
• We have portions of the Aura KB available but they may not be very useful for the current task (grade-level science vs. college level bio).

• Stephen to take lead on definition extractor.

• Send Stephen literature and other material for definition processing.

• What are the research problems?

• Definitions extractor
• Reading rules from text.
• Abductive reasoning
• Procedural escapes for textual matching.