Current Research in Machine Translation: A Reply to Somers PIERRE ISABELLE Canadian Workplace Automation Research Center; Luval, Quebec H7V 2X2, Canada 1. INTRODUCTION The gist of Somers’ paper can be summarized as follows. The classi- cal second-generation (G2) machinetranslationsystem architecture is so flawed with ‘general or specific deficiencies’ that it has not and will not be successful in producing practical MT systems. Most current research efforts are still corrupted by theseerroneous principles. A small number of projects reject the G2 orthodoxy,which makes them better candidates for success. Somerscorrectly characterizes the key feature of the G2 approach as ‘modularity’ along the following axes: 1) different linguistic levels such as morphology, syntax, and semantics are describedseparately (‘strati- ficational’ linguistic models); 2) sourcelanguage, target language(and possibly, contrastive)descriptionsappearin different modules: analysis, generation (and possibly transfer); and 3) linguistic descriptionsarekept separate from processing algorithms. He then proceeds to arguethat this modular designis responsible for a number of fundamental shortcomings. 2. ALLEGED DEFICIENCIES OF G2 SYSTEMS Is Modularity Outdated? The first purported deficiency is that G2 design reflects old-fashioned computationaland linguistic techniques. On the computationalside,Somers sees the G2 architecture as strongly connectedwith the ‘procedural algorithmic strictly-typed programming style’. I fail to see any suchconnection. In fact, just the opposite appears to be true. For example, logic programming turns out to facilitate the development of highly modular descriptions(Isabelle and Macklovitch [4]), which can give rise to reversible G2-typeMT systems (Isabelleet al. 1988,Dymetman et al. 1990). On the linguistic side, Somers associates stratificational models with the outdated transformational-generative (TG) model of grammar.I don’t believethis connection to be anecessary one. Thoughtheparticular number Machine Translation 7: 265-272, 1993. @ 1993 Kluwer Academic Publishers. Printed in the Netherlands.