From DATR to PATR via DUTR - an interface formalism
Markus Duda · OpenGrey (Institut de l'Information Scientifique et Technique) · 1994
There are a number of different formalisms used in the field of computational linguistics like PATR as a unification formalism or DATR as an inheritance based one. One way to partition formalisms is to distinguish between process oriented formalisms and representation oriented formalisms. PATR is process oriented since it is based on unification. The unification operation is binary and closed. This makes it possible to construct arbitrary sequences of unifications to define an algorithm, e.g. the procedure of how to construct a parse tree from a set of input signs. That is why, we can call the unification a procedural operation, and a unification formalism is process oriented. DATR, on the other hand, is a theorem solver. There is no other operation than the built in inference strategy which can not be changed by the user. Thus, the DATR formalism is declarative. Since DATR uses an inheritance mechanism with defaults, it is very simple to define a hierarchy of linguistic objects and it is easy to create new objects from existing ones by similarity. By the help of DATR a linguist can simply model rules with exeptions. DATR can be refered to as representation oriented. It is designed to represent linguistic data. Some unification formalisms make use of an inheritance based type hierarchy. By the help of a default unification operation these formalism are able to express similarity and rules with exeptions as well. The advantage is that a linguist has to cope with just one formalism. The drawbacks are that firstly, there is no such formalism commonly used and secondly, these formalisms are fairly complex. DATR and PATR are very simple and powerful, DATR for representing linguistic data and PATR for processing them. It is useful to combine both formalisms to obtain one tool for linguistic work. (orig.)