Re: RDF: a suitable NLP KB representation (Was: Owning URIs (Was: Yet Another LOD cloud browser))

Sherman Monroe wrote:
> David said:
>  
>
>     I didn't quite express myself clearly. If you were to take the
>     previous sentence ("I didn't quite express myself clearly"), and
>     encode it in RDF, what would you get? It certainly is something
>     that I said about "the thing", the thing being vaguely what I
>     tried to explain before (how do you mint a URI for that?). The
>     point is that using RDF or whatever other non-natural language
>     structured data representation, you cannot practically represent
>     "the things people say about the thing" in the majority of
>     real-life cases. You can only express a very tiny subset of what
>     can be said in natural language. 
>
>
> First off: I began as a NLP researcher seeking the holiest of 
> holy-grails, a method and accompaning knowledge representation 
> formalism with enough semantic rigor to encapsulate any NL statements 
> or expression. What came out of that work was the Cypher transcoder 
> <http://cypher.monrai.com>. When I was first intro'd to the RDF (circa 
> 1999), and when I saw the triple format, it reminded me of predicate 
> calculus (which in my opinion failed the above criteria), and so I 
> turned my noise up at it (and called TimBL a /lunatic/ if I recall), 
> and decided to just work on the NL processing side (i.e. extracting 
> semantics from NL phrase structure) and shelf the knowledge 
> representation side 'til later (i.e. how to serialize the semantics 
> once extracted). Then four years or so later (circa 2003), I made 
> enough headway on the input processing side to turn attention again to 
> the output/knowledge representation side. That's when I was turned on 
> to Frame Semantics, which I immediately praised, it is by far the most 
> expressive and elegant knowledge representation framework for NL I 
> have come across (although, it's been 3 or 4 years since I really 
> looked). In short, frame semantics sees all sentences as a "scene" 
> (like a movie scene) and the nouns all play "roles" in that scene. 
> E.g. a boy eating is involved in a ConsumeFood scene, and the actors 
> are the boy, the utensil he uses, the food, the chair he sits in. So I 
> choose framesemantics as the KB model for Cypher grammar parser output.
Thanks, Sherman, for your story. I had a "history" with Semantic Web 
technologies, too, since 2001. Data on the Web is inevitable. I just 
want to figure out ahead of time what it will actually be like.

> This sent off lightbulbs for me, I went back to RDF, and saw that, low 
> and behold, frames can be represented as RDF, the scene types being 
> classes, a scene instance (i.e. the thing representing a complete 
> sentence) being the subject, the property is the role, and the object 
> is the thing playing that role, e.g:
>
> EatFrame023  rdf:type  mlo:EatFrame
> EatFrame023  mlo:eater  someschema:URIForJohn
> EatFrame023  utensil  someschema:JohnFavoriteSpoon
> EatFrame023  mlo:seatedAt  _:anonChair
> EatFrame023  foaf:location  someschema:JohnsLivingRoom
> EatFrame023  someschema:time  _:01122
> EatFrame023  truthval  "false"^booleanValueType
>
> dbpedia:Heroes(Series) rdf:type dbpedia:TVShow
> dbpedia:Heroes(Series) dbpedia:showtime _:01122
>
> _:01122 rdf:type types:TimeSpan
> _:01122 types:startHour "20"^num:PositiveInteger
> _:01122 types:startMinutes "00"^num:PositiveInteger
> _:01122 types:endHour "21"^num:PositiveInteger
> _:01122 types:endMinutes "00"^num:PositiveInteger
> _:01122 types:timezone "EST"
>
> This says: /No, John didn't eat in a sandwich in a chair in his living 
> room using his favorite spoon, during the TV show Heroes/. Do you 
> still believe RDF is incapable of expressing complex NL statements?
Yes, I still believe. :)

> Second off: Even though RDF (when married with frame semantics) is 
> capable of expressing very complex NL sentences, it was never the 
> intention of the Semantic Web forerunners to create a framework for 
> doing so, and I do not believe that this capacity is nessassary to 
> make RDF valuable. The question RDF answers is fundamentally: /What 
> happens if all the worlds databases (e.g. Oracle, Mysql, etc databases 
> out there) could be directly connected to one another in a large 
> global network, all sharing one massive, distributed schema, and 
> people were able to send queries to that network using a Esperanto for 
> SQL?/ The ability of RDF to represent (not sentences but) rows and 
> columns of any database schema imaginable means it can deliver this 
> vision, and the value tied to it.
And look what happened to Esperanto... After one century, 2 million 
speakers, or 0.025% of the world population.
 
>
>     This affects how people conceptualize and use this medium. If I
>     hear a URI on TV, would I be motivated enough to type it into some
>     browser when what I get back looks like an engineering spec sheet,
>     but worse--with different rows from different sources, forcing me
>     to derive the big picture myself,
>       urn:sdajfdadjfai324829083742983:sherman_monroe
>          name: Sherman Monroe (according to foo.com <http://foo.com>)
>          age: __ (according to bar.com <http://bar.com>)
>          age: ___ (according to bar2.com <http://bar2.com>)
>          nationality: __ (according to baz.com <http://baz.com>)
>          ...
>     rather than, say, a natural language essay that conveys a coherent
>     opinion, or a funny video?
>
>
>
> Then it seems you're still not a convert :) As for me, your example 
> here has very obvious value. Remember what WWW did for humans and the 
> huge revolution that came with giving people access to what other 
> people in the world were saying no matter where in the world they 
> were, and no matter what langauge the host machine spoke natively. The 
> SW is doing that all over again... but for machines this time.

> User empowerment is a large external benefit of the SW, in WWW, 
> webmaster makes assumptions (sometimes rightly, sometimes wrongly) 
> about what data is important and should be shown and how, in SW, user 
> decides for his/herself. Additionally, NL will play a big part of 
> cleaning up the UI so that it doesn't look like an enginerring 
> schematic :) Again, I reference razorbase <http://www.razorbase.com>. 
> Notice the descriptions in the breadcrumbs and descriptions of facets 
> under the 'Your query' link.
Two related thoughts:

At the beginning of the Web, you interacted with the Web by first going 
to a known web site such as your university's site. Then you clicked 
links, saved bookmarks, until you got a number of useful links 
accumulated locally in your browser. That was very congruent with the 
hypertext document paradigm--decentralized, hyperlinking. But then when 
the Web grew too much, we needed search engines. Centralized. Hmm... So, 
we're now building semantic web browsers that are congruent with the 
Semantic Web's paradigm (because if not, you don't get pats on the 
back). Maybe we should start thinking of something ... incongruent? :)

Media are notoriously hard to understand, from what I can understand. If 
we were to say that television was radio but "just" with images, then we 
would be missing something huge. Or that printing was writing but "just" 
much faster. Or that writing was speech "just" recorded on paper. 
Consumer digital cameras are cameras, but just smaller and cheaper and 
faster to develop. Cell phones are phones but just without cords. Etc. 
etc. Is the Data Web the Web just with data? Just for machines? Is the 
difference just that the user can now combine data from several sources? 
How often is that desirable? (Think of your experience today: how often 
would you be willing to pay $1 for RDF from some web page? Daily? 
Weekly? Monthly?) What are the second-order effects?

David

Received on Wednesday, 20 May 2009 05:45:23 UTC