The Vauquois triangle, proposed by the French machine translation pioneer Bernard Vauquois in 1968, is the classic diagram for thinking about how a translation system can work. It is drawn as a triangle: the source language sits at the bottom-left corner, the target language at the bottom-right corner, and the system can take one of three routes between them depending on how high up the triangle it climbs - that is, how deeply it analyzes the meaning of the input before generating output.
The lowest route is direct translation, which works almost word by word using a bilingual dictionary and minimal restructuring. The middle route is transfer, which first parses the source sentence into an abstract structure, applies rules to convert that structure into a target-language structure, and then generates the output. The highest route is interlingua, which analyzes the source sentence all the way into a language-independent representation of its meaning, from which any target language can be generated. As you climb the triangle, you need more analysis but less direct knowledge of the specific language pair, which is why interlingua promised that N languages could be connected with N analyzers and generators rather than N-squared transfer rule sets.
Vauquois built these ideas into the machine translation group he founded at Grenoble in 1960, which developed Russian-to-French and other rule-based systems. His and Christian Boitet’s 1985 Computational Linguistics paper “Automated Translation at Grenoble University” describes how the Grenoble approach used transfer through deep linguistic structures, a concrete realization of the triangle’s middle level.
The triangle remained the dominant mental model of translation for the entire rule-based era and still frames how people describe the field today. Modern neural systems, in a sense, learn their own representation high up the triangle: Google’s multilingual neural translation showed signs of an emergent interlingua-like internal representation, achieving with learned vectors what an earlier generation tried to hand-design.
Why business readers should care: the same three-way choice - shallow and cheap, structured and rule-driven, or deep and general - recurs whenever a company decides how much it should “understand” a problem before automating it, and the history of translation shows that the deepest approach is the most powerful but also the hardest to make work.