NEGRA is part of an interdisciplinary research project, the Sonderforschungsbereich 378 "Ressourcenadaptive kognitive Prozesse" (resource adaptive cognitive processes). The three year project, that has started at the beginning of 1996, is funded by the German Research Foundation (Deutsche Forschungsgemeinschaft). The project is located at the University of the Saarland.
The overall goal of the Sonderforschungsbereich is to study the adaptation of cognitive processes to limited resources. Limitations of cognitive resources under investigation are: the size of the working memory, and time constraints on processing. Resource adaptivity is investigated from several disciplines, such as Computational Linguistics, Computer Science, Psychology and Philosophy. Close interaction between the disciplines is guaranteed by the overall concept of the project, and by specific collaborations between selected disciplines in several parts of the project.
NEGRA is carried out in cooperation of the Department of Computational Linguistics and the Department of Computer Science.
The project addresses the following problems we currently face in natural language processing:
Grammar theories typically model human language competence. But there is a number of phenomena competence grammar fails to give suitable explanations for. Especially word order and rearrangement phenomena of standard word order are hardly to explain in terms of competence only. Another shortcomming is due to the fact that competence grammar is only concerned with the correct sentences of a language. In applications on the contrary, we are interested in real-time processing of sentences occuring in real text or speech. Thus only the most plausible analysis of every sentence occuring is of interest, and not all possible analyses of only the grammatically correct sentences.
Declarative constraint-based grammars, which are state of the art in computational linguistics, provide different types of linguistic information in parallel. Standard parsing strategies operate on a sequential basis. Thus there is necessarily a shortcomming in processing. Linguistic phenomena such as word order variation, discontinuous constituents, and structurally ambiguous constituents cannot efficiently be processed within an entirely sequential approach.
The aim of the project is to develop a new approach on grammar that integrates language competence and performance, and thus overcomes the restrictions imposed by the current view of grammar as competence-based only. The grammar model will mainly aim at linearization phenomena. A new grammar formalism will be developed, that builds upon functor-argument relations. The distance between functor and argument will be expressed by a numeric value, the Bindungsstaerke. Linearization will result from the interaction of various types of constraints, such as valency, topic-focus, syntactic weight of constituents, Bindungsstaerke, and last but not least the size of the working memory.
The formalism will be implemented in Oz. The programming language Oz is specifically well suited for the above mentioned task, as it easily allows for intergration of feature constraints and numeric constraints. The design of the internal contol of Oz also supports expression of cognitively motivated control strategies. Thus implementation of linearization phenomena wrt. to restrictions of the working memory, and parallel processing of various constraints is strongly supported by the programming language.
Thorsten Brants, Denys Duchier, Brigitte Krenn, Martin Müller, Joachim Niehren, Oliver Plaehn, Wojciech Skut, Gert Smolka, Hans Uszkoreit.
Blue Notes for Projects C1, C2, C3, C4, B1, and B4 of the SFB 378.