Another approach I was thinking of is moving a "sliding window" with a length of 3 or 4 words over the texts to be compared. Every list of 3 or 4 words is stored in a hash (with these words concatenated into one as the key and the number of occurences as the value) and thereafter the hashes of both texts are compared to each other; every "match" would give plagiarism-points and every non-match will give originality points: comparing the ratio of the plagiarism-score to the originality-score one can perhaps use this as another metric.

It would be a more abstract metric than your comparison, but it would perhaps be less prone to small deliberate changes or use of synonyms. In that respect one could vary the "sliding window" to look at e.g. first, second and fourth words (skipping the third word) so as to account for substitutions.