This is a lossy algorithm. You will lose information during the translation and reverse-translation steps. Introducing loss into any algorithm obviously increases the difficulty in pulling the clean plaintext out, since it's potentially impossible to pull the clean plaintext out even with the key.
Even so, fairly normal cryptanalysis should apply here. You just need to use a multi-language frequency table. I assume that all languages involved will use a single alphabet? You would drop accents from languages that have them? This is no different than the common step of merging "I" and "J" in ciphers. If you rely on Romance languages, cognates would hurt you. If you go beyond Romance languages, then you're faced again with possibly lossy transliteration, particularly if you mix in Asian or Middle Eastern languages.
This is not dramatically different than the common technique of intentional misspellings. Take a look at Kryptos for a very advanced example that involves the kind of complications you have in mind.
I'd say that your idea makes analysis harder, but not fundamentally different.
When you make this a non-lossy translation, you're describing a homophonic cipher. A book cipher is an example. The idea is that you have different symbols representing the same letter or word. In your case, you're replacing the symbol "road" with either the symbol "rue" or the symbol "camino" or the symbol "lu" for example. You are then applying an alphabetic cipher to the result. This has been used for a several hundred years to prevent frequency analysis. Look at the Great Cipher for a sophisticated version of this approach. Note that as you've describe it, I strongly doubt you'd be able to devise a group of languages that resisted frequency analysis. You'd eventually have to start creating more random symbols (words) until this looked very unlike the original design in my estimation.