Measuring similarity of semi-structured documents with context weights