Classification of Prepositional Senses for Deep Grammar Applications * Lars Hellan Dorothee Beermann Institute for Language and Communication Institute for Language and Communication NTNU NTNU Trondheim, N-7491 Trondheim, N-7491 lars.hellan@hf.ntnu.no dorothee.beermann@hf.ntnu.no * This work is supported by the Norwegian – English Machine Translation project LOGON (http://www.emmtee.net/). Abstract We describe a sense disambiguation tool for prepositional and adverbial relations using Minimal Recursion Semantics as our formal frame. Through defining their basic geometrical properties, we general- ize over temporal and spatial senses to de- rive the main sense distinctions for the most frequent Norwegian prepositional and adverbial expressions. We show how sense distinctions can inform computa- tional grammars and discuss two possible scenarios for enriching HPSG grammars with lexical semantic information: One uses a protégé OWL plugin to create a semantic hierarchy over spatial-temporal senses, with its output to be handed back to the grammar in a post-processing step. The other scenario is the extension of the grammar’s existing type hierarchy with the relational types developed here and the use of Postgres SQL as its lexical module, which allows us to make sense distinctions available only when needed. The present work is conducted as part of the MT project LOGON( http://www.emmtee.net/ ). It also serves as a component of an application in Informa- tion Extraction where lexical sense dis- ambiguation is used to summarize tour descriptions in a language independent format. 1 Introduction Disambiguation and Identification of preposi- tional senses are well-developed components in lexical resources such as Framenet (http://www.icsi.berkeley.edu/~framenet/). Within current grammar engineering, however, little use is made of this type of information. A major reserva- tion concerning the integration of lexical sense dis- ambiguation into medium to large-scale grammars is the increase of complexity that the incorporation of fine-grained semantic distinctions may inflict on these grammars. Given the possible benefits of advanced semantic specifications with respect to applications such as information extraction, ques- tion-answering and machine translation, however, ways around the complexity problem ought to be sought. One possible strategy is to keep the core grammar relatively shallow, and add semantic spe- cification in a 'post-processing' phase. See section 5 for a description of such an approach. Another strategy, more closely linked to the grammar itself, is to define ‘deep’ semantics as a detachable mod- ule of the grammar itself. The latter approach is briefly described in section 6. As a prerequisite for pursuing either course, a reasonably self-contained system of semantic specification needs to be de- fined, in a formalism that can be linked to struc- tures produced by the grammar. The present paper reports on work concerning semantic specification of locative and directional prepositions, developing a system whose general architecture is compatible with any grammars that supports a MRS module. In the present case the system is implemented for an HPSG grammar of Norwegian. The need for a principled, computationally tracea- ble and linguistically wellfounded system of pre- positional senses has been particularly evident within the machine translation community, where the lack of such a system has been up to now a se- rious draw-back, for example for PP-ambiguity management. Also on grammar inherent grounds,