Tuesday, July 05, 2011

Computer source criticism?

TECHNOLOGY WATCH: Computer source criticism?

Algorithm developed by Israeli scholars sheds light on the Bible’s authorship (AP).

As described by this article, this new algorithm, which has been developed by a team led by Prof. Moshe Koppel at Bar-Ilan University, might just be useful. It sounds like a fairly blunt instrument; the source criticism of both the Pentateuch and Isaiah is considerably more complicated than indicated, although this may just be the article simplifying what the program actually did. The designers have clearly put some thought into verifying their results independently (with the experiment on the mixed texts of Ezekiel and Jeremiah) and seem to have gotten encouraging results. I want to see this published in a peer-review journal, to have the raw data looked at by both biblical scholars and computer experts, and to have more tests run on documents whose sources we can verify independently for comparison. But on the basis of this early report I am cautiously optimistic.

The Singularity may be nearer than I thought.

UPDATE: Reader David Clark is skeptical:
I saw your post this morning on computer source criticism. I intended to add a comment to your blog, but it appears you do not allow comments. In any case, software like this make me very suspicious for two reasons.

First, the software tends to know what humans know. That is, they are biased by how the researchers code the software. Thus, if researchers are testing a theory about word choice and authorship, they will necessarily code in information about what a word choice is and how that may relate to authorship. Software like this tends to find what the authors think it will find, no matter how objective and careful they tend to be. Even very subtle biases in a complex computer program behave in non-linear ways, which can have disproportionate effects in the outcome. My point is that even though I think the scholarly consensus on the Pentateuch and Isaiah is correct, I highly doubt this kind of program adds much new insight.

Second, these kinds of software programs have played out already in other religious communities. The best one I can think of is the Mormon faith and the attempt to identify the author(s) of the Book of Mormon. Teams made up of mostly Mormons have written software to identify authorship, as have teams of people not affiliated (or no longer affiliated) with the Mormon faith. Not suprisingly, the software of the former tends to find that the Book of Mormon was not authored by Joseph Smith or his contemporaries, from which they conclude this is evidence of a divine origin. While the latter tend to find evidence for the authorship of the Book of Mormon by Joseph Smith or one of his contemporaries. The bottom line is that computer programs tend to find what people want them to find.
Much depends on whether the software can deduce the sources in documents that were not in the minds of the developers and whose sources we know with confidence on other grounds. If it can do that, I will be impressed.

UPDATE (20 July) More here. Also,this e-mail from reader Ed Kaneen came in on the 11th and I've been meaning to post it, so I will do so while I'm linking to this post.
I'm afraid I've only just picked up your post of 5/7 on Computer Source Criticism, and David Clark's comment. As a former computer scientist, I don't feel the criticism is entirely justified in this case. The original paper offered to the Association of Computational Linguistics Conference is available here: http://www.dershowitz.net/files/unsupervised-decomposition-of-a-document-into-authorial-components.pdf. The method used by Koppel et al. is to first of all group units of text according to the presence of synonyms (as defined by Strong's Concordance), on the assumption that one author will prefer one set of synonyms, and another a different set (this is explained under 'Exploiting Synonym Usage' and they give some examples in section 1.3). This, they suggest, fairly reliably assigns portions of texts to separate authors. However, there are not many such synonyms. Therefore, on the basis of the few units that can be separated in this way, they then use the combination of common words (which occur >5 times in 10, unnamed, biblical books) in these already identified sections as a seed from which to automatically learn a measure for discriminating between the remaining non-identified sections. Their test texts consist of random combinations of two biblical books, primarily Jeremiah and Ezekiel but others too. They do this first at the chapter level, and then at the verse level (i.e. by drawing a random number of chapters/verses from Jeremiah, and then a random number from Ezekiel, and so on – described in section 1.6). Their method seems to be remarkably successful at separating these chapters/verses into their correct original books. They likewise claim in their conclusion a 90% success rate in identifying P and non-P in the Pentateuch (based on Driver).

The point of the above description is that their method is essentially automated, and not dependent on the results they are trying to find. Having said that, of course they will optimise the method to get better results, that is the point isn't it? Nevertheless, I suggest there is at least one problem with both the method and this experiment. With respect to the method, the assumption that there are exactly two authors for a text is a significant restriction – can this be overcome or does the lack of data make the method unreliable for innumerable authors? What results would have been given by applying the method to, say, Ezekiel alone? If the two-author restriction cannot be overcome then it not useful to biblical studies, except perhaps as some measure of confirmation of pre-existing hypotheses. The problem with the experiment is how reliable is the combination of KJV and Strong's Concordance in identifying synonyms, and where does a style of deliberately using synonyms for variety fit into this? In this case however, we should bear in mind that their primary interest is in the novel computational method used, rather than the problem of biblical authorship.

In sum, this is an impressive result, but unsurprisingly relies on the quality of biblical scholarship at its source.