Friday, 18 July 2008

Advertent Curt

Wohoo, I handed in my BA thesis topic today:

From Questions to Queries

Enhancing the Curt System by a Theory and Implementation of Embedded WH-Questions

The Curt System is implemented in Prolog and presented in the last chapter of Blackurn & Bos (2005), a brilliant book about Vincent, who loves Mia, every owner of a hash bar and foot massages. Oh, and computational semantics. And boxers.

Curt is short for Clever use of reasoning tools. Blackburn & Bos use external inferencing tools, a DCG grammar and an emulation of λ-Calculus in Prolog to put together a nice system that is able to parse sentences according to a toy grammar, construct (discourse) models and logic representations of the discourse. They use inferencing to check the discourse for satisfiability and validity.

They develop several variants of the Curt System throughout that chapter, gradually introducing external inferencing, satisfiability checks, validity checks, model building, world and situational knowledge, ontologies and, last but not least, a way to interpret questions in their framework. The last version of the system is codenamed "Helpful Curt" - it can answer basic direct WH-questions by substituting the WH-phrase for an existential quantification over an uninstantiated variable and running a bit of inferencing magic on it in order to find a result.

Advertent Curt

But I found that a bit ad-hoc. Basically, all of the system is a tad ad-hoc (they use Keller Storage for it still!) and it could use some work in order to become really awesome (although it's already quite awesome!). I'd like to show in my thesis that it's possible to introduce a formal account of question semantics into the system. I'll be following Groenendijk & Stokhof's analyses for the most part, also arguing about Karttunen's earlier approach. The main reason for me choosing G&K over Karttunen was that the former use Ty2 to describe their system, whereas Karttunen stays in traditional Montague IL, which would be harder to transfer into the Prolog world.

Ideally, the system should be able to handle the following types of questions:

(1) Vincent knows whether Mia loves Marcellus
(2) Marcellus believes Mia loves Vincent
(3) Every owner of a hash bar knows who likes hash

But why embedded questions?

Almost all accounts of formal question semantics (I know of) treat direct questions as a subtype of embedded questions (Well, except Hamblin, but that's obvious[ly not adequate here]). Once we solve embedded questions, implementing a few transformation steps to translate (4a) into (4b) should pose no problem.

(4 a.) Tell me who owns a hash bar.
(4 b.) Who owns a hash bar?

In this regard I'm taking the very opposite of the more pragmatic approach by Blackburn & Bos, which explicitly handles direct questions only.

I will try to keep the blog up-to-date with progress and ideas :-)

No comments: