Explaining real MT to translators: between compositional semantics and word-for-word

Explaining real MT to translators: between compositional semantics and word-for-word Mikel L. Forcada, Dept. de Llenguatges i Sistemes Informatics, Universitat d'Alacant, E-03071 Alacant, Spain. E-mail: mlf@dlsi.ua.es Abstract Real (i.e., working) machine translation may be presented both as the result of inevitable approximations over an ideal, theoretically motivated model based on the principle of semantic compositionality and as the re- sult of a set of necessary refinements over a very rudimentary word-for-word substitu- tional system. This paper explores the ped- agogical value of presenting real MT as be- ing somewhere in the middle of these two extreme scenarios. I contend that it is possi- ble to reshape the (either optimistic or pes- simistic) expectations of students about real MT by showing students, on the one hand, the approximations, compromises, and sac- rifices necessary to mechanize efficiently a linguistically accurate model, and, on the other hand, the large amount of work needed to improve a word-for-word model so that it produces reasonable translations. This prepares them to learn and appreciate the strategies used to tackle the problem of ma- chine translation. 1 Introduction One of the main problems when teaching MT to translators and linguists is their attitude towards this technology, which may be partly due to the disappointment of high expectations promoted by MT manufacturers. For linguists and translators to be able to understand and value the intrinsic dif- ficulties of MT and the achievements of state-of-the-art systems, it is important to place real MT as being in the middle of two very distinct extreme scenarios, which have to be carefully explained: an ideal model based on an accurate appli- cation of semantic compositionality and a very rudimentary word-for-word, me- chanical, substitutional system (model zero). 2 MT strategies as ap- proximations to an ideal model 2.1 The principle of semantic compositionality Many students think that semantics "studies the meanings of words", but know too well that translating is more than translating words. Semantics has, however, a lot to say about transla- tion: the principle of semantic compo- sitionality (PSC, Radford et al. 1999, 137 p. 359) states that the interpretation (meaning) of a sentence is composition- ally built from the interpretation of its words, following the groupings dictated by its parse tree. The PSC formulation suggests that in- terpretations have a structured nature because they are built by composition of simpler interpretations; it is therefore an ascending or bottom-up formulation (interpretations are built from leaves to root). The PSC is crucial in linguistics because it explains why humans can as- sign interpretation to (i.e., understand) sentences they have never heard or read before. A more general and symmetric ac- count of the PSC (Tellier, 2000), which works not only from sentence to inter- pretation but also in reverse, says that there exist two bidirectional mappings: 1. a mapping between words and ele- mentary interpretations and 2. a mapping between syntactic rules (which indicate how constituents are built from other constituents) and semantic composition rules (which indicate how partial or to- tal …

Content truncated to 500 words