The State of Machine Translation in 1996

In a paper entitled The State of Machine Translation in 1951 [1], Yehoshua Bar-Hillel made it clear that in the absence of methods for resolving " semantical ambiguities " , fully automatic general-purpose MT could only be achieved at the cost of inaccuracy for the " foreseeable future ". Forty-five years later, we are forced to admit that his pessimism has proven entirely justified. Progress in MT research has been painfully slow, and the ultimate goal has kept on receding with each step forward. Traditional rule-based approaches (whether of the linguistics-based or the knowledge-based variety) have resulted in systems that tend to collapse under the weight of their own complexity well before they rise above the clouds of semantical ambiguity. Recent results in corpus-based approaches (whether of the statistical [2] or the example-based variety) indicate that it may soon become possible to develop systems that match the performance of the best rule-based systems with substantially less human labour; but these results do not appear to bring us much closer to high-accuracy MT. Bar-Hillel's pessimism about the prospect for an overall solution to the MT problem was counterbalanced by his confidence that given adequate application strategies, suboptimal solutions could soon be made available. He went so far as to suggest three such strategies. The first was to use MT systems to produce rough translations for the purpose of " scanning the huge printed output [of actual or potential enemies] ". As a tool for information assimilation, MT has indeed proven useful: intelligence services have been using systems like SYSTRAN since the early sixties. Moreover, this kind of application is now becoming a mass-market business under the guise of " browsing tools " for the Internet. For example, Web Translator, a multilingual Netscape-integrated MT system is now available from Globalink for less than $50. No doubt, countless new users will be dazzled both by the machine's proficiency at producing " stammering streams of verbal wreckage " ([4], p.2) and by their own surprising ability to decipher it, at least in part. The second of Bar-Hillel's proposed strategies was to target " restricted language " situations. In 1977, the successful implementation of the METEO system [3], which translates Canadian public weather forecasts, confirmed that MT can indeed work extremely well when applied to simple natural sublanguages. Unfortunately, there are so few situations of this kind on the translation market that METEO remains an …