Patrick Blackburn & Johan Bos (2003): Computational Semantics. Theoria 18(1): 27-45.
is a mix of formal semantics, computational linguistics and automated reasoning. The goal is to convert NL into a formal logic suitable for automatic inference.
Why is it useful?
Because built systems will have applications in many areas such as "information retrieval, information extraction,
dialogue systems, question answering, interpreting controlled languages...". Then it will be important in progress our understanding of how NL works for example will help in understanding the interrelations between syntax, semantics and inference.
It depends on what you want to do, level of detail (linguistics phenomena to represent). They argue that FOL is a good starting point.
How to convert NL to the chosen logical representation?
- Why FOL is a good starting point?
- FOL theorem provers have reached good performances (but still scalability haven't been addressed (see Vampire).
- FOL can represent many interesting phenomena of language like: intensionality (see this page for a description on intension/intentional and other references), temporal relations and plurals (i didn't understand why plurals are not supposed to be easy to represent in FOL, maybe they refers to generics? But from the example it doesn't seem so). The technique used is to add to different types of objects in the domain, for example possible worlds. The problem is that one has to express in FOL how these new objects work, and that is sometime impossible. The paper gives references to other papers discussing this but doesn't give examples.
- But (they say) the main criticism of FOL is its lack of "dynamic potential": with that they mean how to take care of dynamic features of language like references. I'm not sure that their example at page 6 is actually properly done in FOL, i'm not sure that "she is afraid" would be converted into afraid(x), but basically the problem is an unbound variable and that is stated clearly. The conclusion is that FOL has many faces (for example, Dynamic Predicate Logic and Discourse Representation Theory) and each face can be better at something. So, if we want to use FOL we have to be willing to use the appropriate face instead of the default face.
They discuss two approaches: unification-based approaches and lambda-based approaches. The choice here is independent from the logical representation chosen before. They prefer the lambda-based but only because more clean. I should scan through 447 notes to see more details.
How to deal with the ambiguities in NL?
The most common approach is underspecification
, that means that you build the formal representation leaving ambiguitis and then you resolve them in the logical representation. See the DORIS
system for an implemented system (it appears not to work as of today).
they briefly discuss theorem provers and model builders.
In the conclusion they refer to the possibility of using directly in the inference machinery underspecified formulas.
- 06 Dec 2006
They mention a tradeoff between the demands of expressivity and inferencial effectiveness. There is no such a tradeoff, Len proposed this thought experiment:
- take a more expressive system EL with its reasoning engine, and a less expressive system FOL.
- given a query as input, first check if parsable by the FOL system, if yes call the reasoner of FOL to answer it.
- if not use the reasoner for EL to answer the question.
So, there is no substantial loss in inferencial effectiveness even in a more expressive language.
One certain advantage of using FOL is the availability of many theorem provers.
There is a tradeoff between intensionality and extensionality, in the limit one can add more and more to the domain of discourse and what before
was intensional become extensional. Len mentioned that type theory goes in this direction. See Shalom Lappin's type theory
Problems of davidsonian approaches to representation of temporal relations: negation and references of complex events. Examples given: "It didn't rain for 3 days, the crops was damaged by that." and "The situation of each superpower menacing the other with nuclear weapons lasted for 2 decades.".
- 08 Dec 2006