Defining the Lexical Component in Interlinguas

As discussed by Dorr and Voss (1993, 1994), machine translation (MT) theory has not addressed the issues surrounding how the interlingua (IL) of a MT system should be defined or evaluated. This has a direct bearing on the decisions developers make with respect to the construction of a lexicon for MT. We view the IL as two distinct components: the declarative portion, which we call the "Lexical Component;" and the procedural portion, which we call the "Pivot-Form Component." The former component is a collection of entries from each natural language lexicon of the MT system. The latter component is the set of algorithms used to compose and decompose the full IL pivot form. 2 The focus of this paper is on the definition of representations in the Lexical Component. We note, however, that the Lexical Component must be be tested in conjunction with the algorithms of the Pivot-Form Component. In particular, when IL representations are defined in the lexicon, 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. Interlocking problems may arise as the developers of an interlingua start to build the lexical IL forms for the lexicon and then attempt to write pivot-form algorithms that are compatible with the lexical IL fomls. During algorithm development, they must then take into account the range of sentential contexts where each lexical item may appear; this often forces revisions to the IL forms in the lexicon. We briefly examine the interlinguas in the lexicons of five current IL-based approaches to MT. Even with this limited review, the overall Rucllng is clear: no consensus exists ~mong MT researchers for delqnlng (the equivalent of) the Lexical Component of an with respect to the levels of representation in an MT system. We provide a sketch of a tool, ILustrate, 3 currently under development for the design and evaluation of different lexical representations for interlingual MT.