3 ms·

Question 3 was the most interesting but data on parsing remarkably incomplete. As far as I can tell, we have only lists of possible ways to break down the data

by Jd 16y ago

Question 3 was the most interesting but data on parsing remarkably incomplete. As far as I can tell, we have only lists of possible ways to break down the data without any explanation of how or why one possible way is preferred to another. Case in point (1):How it decides to treat "Treasure Island" as a proper noun. We see only "modifies(Treasure, Island)" -- indicating that it treats treasure and adjective modifying island, then suddenly in the semantic assumption phase they are treated as a compound. Case in point (2). We are given: island(Treasure Island) location(Treasure Island) resort(Treasure Island) book(Treasure Island) movie(Treasure Island) I assume what he is giving us is method names written in Java with "Treasure Island" as the single argument that return a value indicating the likelihood that "Treasure Island" is what the method name refers to. This is extraordinarily interesting. However, it is not at all clear which methods are chosen and why, if they are run in some sort of sequence or simultaneously, etc . Case in point (3) : "Builds different semantic queries based on phrases, keywords and semantic assumptions." This is very vague but indicates that Watson generate a set of queries which it runs against its own internal search engine ranking answers presumably based on the quality of the initial search and the confidence of the answer. Would be very very cool to have an example. All in all, wets the appetite but leaves one wishing for more hearty fare (or a job at IBM!).

ppod 16y ago
A while ago I submitted a link to a blog post( http://bit.ly/igJeRB http://bit.ly/igJeRB ) I did that goes into the system in a bit more depth, based on a paper IBM published - there's a link to the paper (open access) there too. For your cases, from what I got from reading papers: (1) You could spend two weeks only reading papers on noun-compound semantics. Try a google scholar search just to get an idea of the volume of research. A simple technique to test how idiomatic a phrase is would be a Bayesian type test to see how many times "Treasure Island" occurs in a corpus divided by how many times "Treasure X" and "X Island" occurs. In this case the capitalization probably cues it to look up Treasure Island in Freebase. Interesting thought actually - do the contestants also get the question as text? I think they do, so they get capitalization. (2)I would be pretty sure these are not Java methods, I'd say they are logical predicates representing the fact that 'Treasure Island' is returned as being a member of the set of things indicated by the predicate, as returned either by syntactic processing (island) or from the knowledge bases (WordNet, Yago, Freebase, dbpedia) (3) there isn't a worked example in the paper, but my idea of this is that it's basically Watson's way of figuring out what queries to type into it's unstructured text corpus (they have a corpus of web snippets indexed with lucene)
- natrius 16y ago
  For more on (2), see http://en.wikipedia.org/wiki/First-order_logic http://en.wikipedia.org/wiki/First-order_logic
lmkg 16y ago
> I assume what he is giving us is method names written in Java with "Treasure Island" as the single argument that return a value indicating the likelihood that "Treasure Island" is what the method name refers to. The sample queries struck me as Prolog rather than Java, even before they mentioned it. Using Prolog would allow testing the combination of various propositions in addition to the individual values. This would imply that they derive a list of possible assertions, then try combinations of those assertions and see what values of X will match those assertions. How they avoid combinatorial explosion when testing all those assertion combinations, I have no idea, but it must involve ranking and pruning.

by Jd 16y ago

ppod 16y ago

A while ago I submitted a link to a blog post( http://bit.ly/igJeRB http://bit.ly/igJeRB ) I did that goes into the system in a bit more depth, based on a paper IBM published - there's a link to the paper (open access) there too. For your cases, from what I got from reading papers: (1) You could spend two weeks only reading papers on noun-compound semantics. Try a google scholar search just to get an idea of the volume of research. A simple technique to test how idiomatic a phrase is would be a Bayesian type test to see how many times "Treasure Island" occurs in a corpus divided by how many times "Treasure X" and "X Island" occurs. In this case the capitalization probably cues it to look up Treasure Island in Freebase. Interesting thought actually - do the contestants also get the question as text? I think they do, so they get capitalization. (2)I would be pretty sure these are not Java methods, I'd say they are logical predicates representing the fact that 'Treasure Island' is returned as being a member of the set of things indicated by the predicate, as returned either by syntactic processing (island) or from the knowledge bases (WordNet, Yago, Freebase, dbpedia) (3) there isn't a worked example in the paper, but my idea of this is that it's basically Watson's way of figuring out what queries to type into it's unstructured text corpus (they have a corpus of web snippets indexed with lucene)

natrius 16y ago
For more on (2), see http://en.wikipedia.org/wiki/First-order_logic http://en.wikipedia.org/wiki/First-order_logic

lmkg 16y ago

> I assume what he is giving us is method names written in Java with "Treasure Island" as the single argument that return a value indicating the likelihood that "Treasure Island" is what the method name refers to. The sample queries struck me as Prolog rather than Java, even before they mentioned it. Using Prolog would allow testing the combination of various propositions in addition to the individual values. This would imply that they derive a list of possible assertions, then try combinations of those assertions and see what values of X will match those assertions. How they avoid combinatorial explosion when testing all those assertion combinations, I have no idea, but it must involve ranking and pruning.