If we have n languages (where n can be 100 or more), full translation matrix need n^2 models, where translation to intermediate language (IL) and from IL – only 2*n.
IL must have clear syntax rules, be unambiguous, be like contex-free computer languages rather than natural language. Translation to IL need Word -sense disambiguation (WSD ) (for example rock and drop in ML have specifed sense, not lexem)
One candidate to IL was UNL (Universal Networking Language) but was patented and is abandoned. Other project is Abstract Wikipedia and Wikidata lexems, is in the early stages.
Goal: don’t use too many resources for predefined models for many languages and translation available offline in home desktop computer .
The main advantage of English as the universal intermediate language is there’s a lot of data available for English.