The Case for a MT Developers’ Tool with a Two-Component View of the Interlingua

The Case for a MT Developers' Tool with a Two-Component View of the Interlingua1 Bonnie Dorr and Clare R. Voss Department of. Computer Science University of Maryland College Park, MD 20742 {bonnie,voss}@cs.umd.edu Abstract The interlingua (IL) in machine translation (MT) systems can be defined in terms of two components: (i) "lexical IL forms" within language-specific lexicons where each lexical entry has associated with it one or more lexical representations, and (ii) algorithms for creating and decom- posing the instantiated "pivot" representation. Within this framework, we examine five different approaches to the level of representation for the lexical IL forms and then discuss a tool, ILustrate,2 we are building to develop and evaluate different IL representations coupled with their correspond- ing translation algorithms. 1 Introduction As discussed by Dorr and Voss (1993), MT theory has not yet addressed the issues surrounding how the interlingua (IL) in machine translation should be defined or evaluated. This paper provides a framework for addressing such issues. We take the IL in machine translation to be defined implicitly in two distinct ways: (i) declaratively and (ii) procedurally. We refer to the declarative portion of the IL as the "Lexical Component," consisting of the collection of IL entries from each natural language lexicon of an MT system. We refer to the procedural portion of the IL as the "Pivot-Form Component," consisting of the algorithms for creating and decomposing the full IL pivot form. 1.1 Interlocking of the IL Components At IL definition time, the decisions concerning the two IL components are frequently interlocking, i.e., a change to one component drastically affects the functionality of the other, and vice versa. At MT run-time, by contrast, the relation between the components is fixed and one-way, with the pivot-form algorithms accessing the lexical IL forms. We briefly touch on an example of the interlocking problem here. Consider the English words about, by, down, in, off, on, out, over, under, and up. These appear in two syntactic constructions (Lindner (1981)): the verb-particle construction (VPC, a complex with a verbal element and a particle) and the verb-prepositional phrase construction (VPP, a verb plus PP that may be an argument or an adjunct to the verb). These two construc- tions are illustrated here as (1) and (2), respectively: 1 This research was supported, in part, by the Army Research Office under contract DAAL03-91-C-0034 through Batelle Corporation, by the National Science Foundation under grant IRI-9120788 and NYI IRI-9357731, and by the Army Research Institute under contract MDA-903-92-R-0035 through Microelectronics and Design,Inc. 2 The acronym ILustrate stands for InterLingua Users' Support Tool, a Research And Testing Environment. We expect that the users of the tool will be MT researchers, MT system developers, and MT system evaluators. 40 (1) a. The intruder shot out the light. (2) a. The dog shot out the door. b. The troops ran down the rations. b. The troops ran down the hill. c. They looked up the address. c. They looked up the chimney. Note that a parser, using syntax …

Content truncated to 500 words