Gripla - 2020, Page 29

GRIPLA28
Since this is the case, we could certainly argue a priori that, when
comparing Íslendingasögur documents, our word frequencies are more
significant since the total range of possible words is narrower (effectively
reducing the dimensionality of the feature space). But proving such a
claim would require a methodological study extending Eder’s research
into the Old Norse field, and such a study still remains to be conducted
(and is much desired). Nevertheless, it is important to note that Eder’s
goal, and the goal of many stylometricists to whom Eder was responding,
was unequivocal authorship attribution. In order to appear as a correct at-
tribution in his study, the stylometric classifier would have to determine
the correct author. In practice, textual researchers are (or perhaps should
be) seldom after such strong classifications. Rather, in situations where
antiquity has afforded us with a limited set of evidence, we must resort
to fuzzier probabilities. Given this, it is worthwhile to point out a similar
experiment conducted by Burrows which found that stylometry could rank
the correct author among a list of top candidates with documents as short
as 150 words.62
Various factors lead us to believe that the results of our investigation
are not based on random chance, but rather genuinely speak to the relation-
ship between Ljósvetninga saga A and C. First, the investigation is rather
simple, targeting a small number of texts, two of which are almost the
same. This means that the dimensionality of the problem is low, which is
helpful. If we were exploring a corpus of hundreds of small texts (as Eder
was), the dimensionality of the problem would be much larger, increasing
the likelihood that the significance of word frequencies would get lost in
the void of an excessive feature space. The documents have also been heav-
ily reviewed by the authors for consistency, which is not the case for many
textual corpora in stylometric literature. Finally and most importantly,
in what follows we conduct a series of tests with different setups and at
every stage the overall pattern of the results is always the same. This is a
good sign, since it indicates that the overall relationship between the docu-
ments (that the C-redaction is most internally consistent with the parallel
62 See in particular Table 3 in John Burrows, “‘Delta’”: A Measure of Stylistic Difference and
a Guide to Likely Authorship,” Literary and Linguistic computing 17:3 (2002): 275. In this
article John Burrows was working with the original delta metric which he devised here, and
it should further be pointed out that metrics have improved since that time.