In Estonian, "in London" is translated as "Londonis". Now if we want to apply natural language understanding to this, then should I tokenize "London" and "is" separately? Or should I lemmatize "Londonis" to "London"? Ofcourse both approaches might be possible, but I'm asking for the rational and recommended approach. The issue with words with suffixes is that the sequence tagging statistical model for named entity recognition is not detecting words with suffixes. And I don't speak Estonian.