Lexicon-Driven Machine Translation

Machine Translation (MT) systems have historically relied upon explicit grammars in order to analyze the source text and reproduce it in the target language. In this paper, we argue for a style of MT in which the focus of processing is at the level of the lexicon, rather than the grammar. This approach to translation allows an analyzer to map source sentences into an interlingual form, which then can be mapped (perhaps after intermediate inferencing steps) back into target sentence(s) which are paraphrase-equivalent to the original. Advantages of the approach include: 1) the possibility for different paraphrases of the original; 2) the capability for multi-sentence expression of the original when no single word (e.g., a verb) exists in the target language which spans the same meaning complex as a word in the source; 3) a uniform approach to word sense disambiguation and anaphoric reference resolution; and, most importantly, 4) the possibility for robust handling of ungrammatical and ellipsed source text. 1. 0 Introduction: Lexicon-Driven Machine Translation Systems designed for the Machine Translation (MT) of texts between languages have traditionally relied upon explicit grammars of both the source and target languages, in order to analyze the source text and produce well-formed target-language sentences ( c f . , e.g., [Tuck84]). Grammar-based systems have been reasonably successful at production-quality MT. The nature of grammatically driven processing leads to certain problems, however. First of all, an explicit grammar tends to make the system’s computation excessively topdown. Thus, it is usually not particularly robust under deviant ( e .g . , ungrammatical, telegraphic or ellipsed) input. Moreover, explicit-grammar approaches tend to be overly concerned with the form of the language, rather than its content. Issues of preservation of meaning between source and target texts tend to get downgraded. In this paper, we shall argue for an alternative style of MT in which the focus of processing for both input and output texts is at the level of the lexicon, i . e . , the words and phrases of a language, rather than its grammar. No extensive experimentation in MT has been performed within t h i s paradigm ( but see, for example, [Wile81, Lyti84]). We shall suggest, however, that the approach very naturally allows for meaning-preserving MT, and provides solutions for di f f i c u l t problems such as word-sense disambiguation, anaphora resolution, the need for circumlocution when lexical equivalents of source words are not available, etc. Language analysis, in particular, has a very strong bottom-up nature in the approach to be described. Such analyzers [e . g . , Ries75, Birn81, Cull84, Dyer84] tend to produce fragmented meaning structures for ungrammatical or ellipsed inputs. Thus, there is the possibility for diagnosis of the fragments, in order to determine a reasonable reading of the input [e . g . , Boot85a, Boot85b]. We illustrate the approach with a toy MT system which translates Ukrainian texts into English. The overall methodology for designing language interfaces in this paradigm is detailed in [Cull85]. * The research described here was performed while this author was visiting the EE&CS Department, Princeton University, Princeton NJJ08544.