Applying edge-counting semantic similarities to link discovery: scalability and accuracy