Graphology

Yesterday I was moved to ask Bing about graphology, which took me to a Wikipedia article which tells me that it is pretty much rubbish, on a par with reading palms and reading horoscopes. See reference 2.

I was not so sure, thinking that some handwriting was distinctly feminine and that, generalising, it did not seem unreasonable that one could glean something about a person from their handwriting. And that handwriting could generate (as well as reflect) emotions in the same sort of way as music, or indeed as art writing, that is to say calligraphy, the sort much used in a religious context by Muslims. Not to mention the likes of Eric Gill, snapped above, more into font design than writing. Or the Book of Kells, at the boundary between writing and painting, of reference 3.

I then turned up one of the many references in the Wikipedia article, the present reference 1, an article about using neural networks to determine the gender of the writer of a piece of handwriting. That is to say, with a computer. With the result that they get it right about 75% of the time - which, in the absence of any statistical commentary, accessible or otherwise, strikes me as significant. It can be done.

The method is fairly stripped down, in that it looks at small patches on images of pages of handwriting, looking for features of that handwriting without regard to what the letters are or what words they are in. So a lot of information, some quite possibly relevant, is being discarded. Should one, at the margin, be allowed to use the fact, if indeed it were a fact, that the vocabulary of men is statistically different from that of women?

After training, the algorithm scores each patch. It then scores the page according to one of two rules. One, the gender of the page is the gender of the majority of patches. Two, the gender of the page is determined by the average of the patch scores. From where I associated to the scoring rules in the ongoing presidential election in the US. Didn't seem to make that much difference here.

Most of the time, the accuracy of gender assignment at page level is 75% or more, rather better than the average human. Don't know about expert humans. With results varying according to the various permutations of training and testing languages, that is to say either Hebrew or English. Given that one is not interested in either letters or words, training in, for example, Hebrew and testing in English does not make as much difference as one might have thought.

It will be interesting to see the results of follow-up work. Perhaps one can do age or race as well?

PS: asking Bing about this Netanyahu takes me to a Wikipedia entry which says: 'Nathan S. Netanyahu (Hebrew: נָתָן נְתַנְיָהוּ‎; born 28 November 1951) is an Israeli computer scientist, a professor of computer science at Bar-Ilan University. Netanyahu is the son of mathematician Elisha Netanyahu and Supreme Court of Israel justice Shoshana Netanyahu, the nephew of historian Benzion Netanyahu, and the cousin of Prime Minister of Israel Benjamin Netanyahu'. I had thought that this last was a paratrooper turned politician, with the style of the older politician built from that of the younger soldier, that is to say of the barrack room. Which is now clearly wrong: he comes from a serious, educated family and had quite a decent education himself. Born in Israel, raised in Israel and the US.

References

Reference 1: Handwriting-Based Gender Classification Using End-to-End Deep Neural Networks - Evyatar Illouz, Eli (Omid) David, Nathan S. Netanyahu  – 2018.

Reference 2: https://en.wikipedia.org/wiki/Graphology. 'Graphology is the analysis of the physical characteristics and patterns of handwriting with attempt to identify the writer, indicate the psychological state at the time of writing, or evaluate personality characteristics...'.

Reference 3: https://en.wikipedia.org/wiki/Book_of_Kells.

Reference 4: https://psmv3.blogspot.com/2018/08/gill.html. One of a number of previous notices of Eric Gill.