Difference between revisions of "Vulcan/SystemPrototype"
From Knowitall
(→Status) |
(→Status) |
||
Line 32: | Line 32: | ||
<blockquote> | <blockquote> | ||
* The system fails on the iron nail example. | * The system fails on the iron nail example. | ||
− | * Both "iron nail" and "plastic cup" get same high probability score (0.94). <b> | + | * Both "iron nail" and "plastic cup" get same high probability score (0.94). <b>Assigning smaller weights on the tuple generation rules fixes this problem. Previously, these rules had a infinite weight.</b>. |
− | * | + | *<b><span style="background-color:yellow">I need to better understand how the MLN scoring works in general and how Tuffy implements the inference.</span></b> |
+ | *<b><span style="background-color:yellow">So why having</span></b> | ||
+ | |||
</blockquote> | </blockquote> | ||
Revision as of 18:57, 3 September 2013
Overview
The prototype is designed to work on three questions. We want the system to output the following:
- Score for the input proposition.
- New facts inferred.
- Facts and rules used in scoring.
Status
- Ran Tuffy on three example questions. It failed on one question.
- Hand generated the input evidence for the propositions (one correct and one incorrect) for three questions.
- Hand generated the MLN rules based on Stephen's human-readable rules.The MLN rules can be found here.
- Ran Tuffy to obtain the inference probabilities on the propositions.
- The system also outputs:
- All inferred facts along with their probabilities.
- All rules that are reachable from the query fact. i.e., Clauses in the MLN that are relevant to the inference of the query fact.
- Does it work?
- Tuffy gets it right for 2 out of 3 questions. i.e., it assigns higher probabilities for the correct proposition.
- Facts inferred by larger number of steps have a lower score compared to those inferred by a smaller number of steps.
- Why does it fail on the one question?
- The system fails on the iron nail example.
- Both "iron nail" and "plastic cup" get same high probability score (0.94). Assigning smaller weights on the tuple generation rules fixes this problem. Previously, these rules had a infinite weight..
- I need to better understand how the MLN scoring works in general and how Tuffy implements the inference.
- So why having
- What diagnostics do we NOT have?
- Connections between the clauses in the MLN.
- A reconstruction/visualization of the MLN network. Working with Tuffy developers on this.
- What next?
- Figure out why iron nail fails.
- Use automatically found evidence i.e., actual open ie tuples instead of hand-generated evidence.
- Figure out what mismatches exist between required and available knowledge.
- Improve diagnostics output.
- What does this exercise suggest?
- Need to figure out how the weights on the MLN rules and evidence are use. [I assigned them arbitrarily for this round.]
- Use predicates with small arity. For example, avoid writing rules entire nested tuples as predicates.
- The only reason we'd need a nested tuple is for the purpose of computing the score. For now we can compute this from the score of its components: Score(nested_tuple) = Score(top tuple) * Score (nested).