Just like all the scoring we do, it's really based on how humans have scored the item. So our computers measure the text and it looks at how well the content of the sentence covers the content of the text as a whole, so is looking at the meaning of both of them and how similar they are. In addition we're also looking to make sure that it's a real English sentence and not just a string of keywords, that the words follow each other in the way the usual English sentence would.
