Some phrases tend to appear more frequently in the document, and not all users build their term bases on the fly. What I'm suggesting here is more or less what AutoSuggest is already doing, but simply using this information to automatically build a termbase. Each term added automatically could have a special field indicating it was added automatically, but I'm pretty sure having an extra asset (termbase) once the translation is complete with no effort would have a great value. Even if not all the entries would be correct (I'm guessing we could have an option to confirm/reject suggested automatic termbase entry one by one or not).