Token-based minimal recursion semantics

I will describe a new computational approach to natural language semantics. I call it Token-based minimal recursion semantics (TBMRS). TBMRS can be seen as a version of Minimal recursion semantics (MRS, Copestake et al 1999), which is way of structuring semantic representations. Both the original and the present versions of MRS have been developed as integrated into a Head-Driven Phrase Structure Grammar (HPSG). However, a semantics of this kind could be supported by various formal grammars. The original motivation behind TBMRS has been to implement the philosopher Donald Davidson’s (1984) so-called paratactic treatment of the semantics for intensional constructions. As far as I know, this is the first attempt of turning Davidson’s proposal into a computational model. Here, I will only mention briefly what the “paratactic” semantics is all about. I’ve developed a TBMRS syntax and semantics for some English constructions, mainly intensional ones (Dahllöf 2001a). This grammar has been implemented with the help of my own system, Prolog-embedding typed feature structure grammar (Dahllöf 2001b, 2001c). A demo is available over the WWW (at http://stp.ling.uu.se/ ̃matsd/tbmrs1/). TBMRS is a formal and methodological framework. It’s not a specific theory. It can be combined with various accounts of specific details. Minimal recursion semantics, or MRS, is mainly a way of organizing semantic representations. The MRS method can be applied to almost any kind of underlying formal semantic framework. The basic idea behind MRS is to use representations with a flat syntax. Syntactic embedding is eliminated. The elementary units of the representations are called elementary predications (EPs). The EPs roughly correspond to the basic semantic units of ordinary logic. The MRS representations are sets or lists of EPs. Each EP is associated with a so-called handle, an identifying label. Some EPs, for instance, quantifiers, carry features which take handle values. This allows feature-handle connections to indicate operator scope. So, syntactic embedding is not needed for that purpose, as in ordinary predicate calculus. The “ordinary” predicate EPs carry a handle value, as all EPs do. Another feature identifies the predicate, i.e. the lexeme. Additional features correspond to the logical arguments of the predicate. Their number and their kinds depend on which predicate we’re dealing with. Apart from handle and lexeme features, generalized quantifiers involve a bound variable, a restriction and a body. The