[Developers] Constructing MQL queries automatically

[John Bäckstrand]

Ah, yes, question answering. I did my thesis (which isn't complete,
btw, ahem) in "question classification" which is just a sub-set of
question answering. The basic problems boils down to understanding
natural language, which is yet largely unsolved. There are however a
few approaches to "classifying" questions, more specifically getting
the "answer type", ie. what kind of answer is expected. I basically
tried an approach where I focused statistically on verbs, and used
wordnet to try and find links to the answer type. I found out that was
all wrong, because the type of question is largely determined by the
other words, such as "what is", "where is", "who is" and so on. The
nouns are most of the time just qualifying the type of question a bit
further.

A search on citeseerX for "question answering" or "question
classification" shows a lot of hits.

On Fri, Nov 7, 2008 at 18:58, Shawn Simister <narphorium at gmail.com> wrote:
> What you're describing is one of the major challenges in question answering:
> how to map sentence structure onto data schema. That's what makes question
> answering a difficult problem in AI. All the approaches that I've seen use
> an explicit list of rules to map questions to queries.
>
> If you were to do it automatically in Freebase you would search for types
> and properties by name like this:
>
> [{
>   "id" : null,
>    "type" : "/type/type",
>    "name" : "film"
> }]
>
> [{
>   "id" : null,
>    "type" : "/type/property",
>    "name~=" : "starring"
> }]
>
> But, these queries still depend on Freebase users to create type & property
> names that fit your application. I doubt that this would be sufficient to
> get usable results in your application.
>
> A much more interesting solution would be to create your own set of Freebase
> types to explicitly map types to nouns and properties to verbs. This would
> allow anyone from the Freebase community to contribute mapping to your
> application and improve the scope of its question answering abilities. If
> you're interesting in that approach I'd be happy to help you out.
>
> Shawn
>
> Vanessa Lopez wrote:
>
> Hi,
> I am trying to develop a Question Answering application on the top of
> Freebase.  However, I have some doubts that maybe any of you have experience
> with ..
> For instance, from the example query : George Lucas films starring Harrison
> Ford
> [
>   {
>     "directed_by" : "George Lucas",
>     "name" : null,
>     "starring" : [
>       {
>         "actor" : "Harrison Ford"
>       }
>     ],
>     "type" : "/film/film"
>   }
> ]
>
> How it is possible to obtain from the keywords "George Lucas", "film",
> "starring" and "Harrison Ford"  that the type we want is "/film/film" and
> the relation we want is "directed_by" automatically?
> For instance, the types of film(s) { "name" : "film(s)", "type" : null} are:
> /music/albums, /common/topic, /type/domain , /freebase/domain_profile, and
> we have 96 ids  { "name" : "film(s)", "id" :null}, being /film/film one of
> them ...
> Also, we can think of other MQL queries, like {"type":"/film/actor",
> "name":"Harrison Ford", film: null, ...} that in principle could have been
> valid too ...
> Not to mention about "starring" ...
> In other words, it does not seem to be straight-forward to translate this
> query into MQL without having a lot of knowledge about topics, properties,
> etc in Freebase ...    Is there anywhere to automatically do this
> translation? does any of you encounter this problem before while developing
> an application?
> Many thanks!
> Vanessa
>
> ________________________________
> _______________________________________________
> Developers mailing list
> Developers at freebase.com
> http://lists.freebase.com/mailman/listinfo/developers
>
>
> _______________________________________________
> Developers mailing list
> Developers at freebase.com
> http://lists.freebase.com/mailman/listinfo/developers
>
>



-- 
John Bäckstrand