If an NLP system processes a text containing proper nouns like names, trade marks, etc. without knowing anything about the language (ie no lexicon), is it possible to recognise them?
Essentially you would need to train a named entity recognizer (NER) to recognize the names out of the common words.
There are many works that try to use a similar language to the language in question as a pivot to train a full NER model (for example, Cheap Translation for Cross-Lingual Named Entity Recognition).
Your task might be slightly simpler, as you are interested only in one class: whether it is a named entity or not. But in general it is very similar to this setup.