How to Overcome Translation Mismatches - An Inference Driven Mapping between Meaning Representations

This paper deals with issues that a bidirectional GermanRussian machine translation system faces when the meaning of spatial prepositions in these languages does not line up. A uniform representation language is used to define the meaning of spatial prepositions in a language independent way. This formal language makes it possible to compare monolingual meaning representations and allows for the recognition and resolution of translation mismatches. 0. Introduction The main problem for multi-lingual machine translation (MT) is that "there is no hope of constructing a language [interlingua] which is such that translations in all languages map onto a single abstract representation in IL" ([KAY 1991], p. 78). Even if the ambitions of the IL approach are confined to the design of a meaning representation mediating between two languages there are many translation mappings which cannot share the same representation. Because of translation mismatches intra-lingual mappings between IL representations are often inevitable. There are several ways to solve translation mismatches. [KAMEYAMA 1991] proposes to add or to drop information depending on whether the target language (TL) renders the expression more precise or allows ambiguity not allowed in the source language (SL). The IL approach taken by [BARNETT 1991] appeals to a monolingual language description where translation mismatches are viewed as generation problems. Here the generator includes techniques that assign preferences to a certain way of expressing an assertion and makes use of SL information to guide this assignment. This paper contributes to the issue of how the translation mapping between non-identical meaning representations can be achieved in an efficient way by preserving languagepair independence in the lexicon. I sketch an approach of an inference driven mapping between monolingual meaning representations and will apply it to the problem of translating between German and Russian spatial prepositions, which are a well-known source of a wide range of translation mismatches. For the locative prepositions considered I propose meaning representations which express the spatial relation between two concrete objects. The representation language for spatial prepositions is based on tools from settheoretic topology, i.e. objects are represented as sets of spatial points and the relations between them are expressed in terms of inclusion, overlap and connectedness. The proposed meaning representations for the German and Russian locative prepositions can be used to infer whether the meaning of the target language preposition is identical with, more general or more specific than that of the source language, These correspondences between SL and TL representations are established by use of identity, superset and subset relations between the sets of spatial points expressed in the prepositions’ meaning representations. Following basically the ideas of the "Translation-by-Negotiation-Approach" proposed by [KAY 1991] my approach makes explicit how translation mismatches can be recognized and consequently resolved by a "negotiator" which operates on language independent knowledge. First, I give a survey of some of the translation problems we are faced with. Second, I present the formal representa: tion language and focus more specifically on issues that concern the meaning representation of spatial expressions in the lexicon. Then, I show how the correspondence between