Oui Oui baguette

Experimenting shapes measurements and comparing handwritting from two different manuscripts written by the same hand

Alexandre Mourlon ·

Abstract

The main idea is simple: the paleographer has the superpower of being able to say things about a manuscript just by looking at it. For example, they can identify its period, its place of origin, and sometimes even its copyist.

I have two manuscripts from the Early Modern period. One is a historical chronicle written in French by Pierre-François Barberot, while the other is a collection of literary pieces written in a mixture of French and the local dialect. There is no signature on the second manuscript, only an ex libris belonging to Pierre-François Barberot. Still, the handwriting is very similar. Therefore, the idea was to take measurements and make comparisons between the two manuscripts.

Result

Heatmap of the cosine similarity of Ms 1007 compared to Ms 1054
Figure 1. Heatmap of the cosine similarity of Ms 1007 related to Ms 1054.

This is unreadable, so please, feel free to access the link below

View the complete visual comparison table

Average similarity for each word
Figure 2. Average similarity for each word .
Heatmap of the cosine similarity of Ms 1007 compared to Ms 1054
Figure 3. Heatmap of the cosine similarity of Ms 1007 related to Ms 1054.
Same-scribe letter variability vs. cross-manuscript similarity
Figure 4. Same-scribe letter variability vs. cross-manuscript similarity.
Establishing an Optimal Similarity Threshold
Figure 5. Establishing an Optimal Similarity Threshold.

The shapes

The full data is in All_Data.zip.

Word (Letter)Mss 1007Mss 1054
Frankendal (F)DinosaureDinosaure
Description (r)Dinosaure
Frankendal (a)DinosaureDinosaure

How it was made

The data comes from two manuscripts held at the Besançon Municipal Library; the first Histoire de la Franche-Comté de Bourgogne et en particulier de la cité royale de Besançon, capitale des Séquanois, composée par le R. P. Léopold Prost, de la Compagnie de Jésus, et continuée par le sr Pierre-François Barberot, avocat en Parlement, citoyen de Besançon (Ms 1007) and the second is Plusieurs dialogues, tant en françois qu'en patois, sur Besançon, receullis (sic) audit lieu l'an de grâce 1715 (Ms 1054). The idea is to compare the handwriting of the two

The first step was to segment the characters. This was done simply by taking screenshots of the digitized documents. I decided, entirely arbitrarily, to select 7 words from each manuscript,64 letters per manuscript,for a total of 128 characters. I chose the same words from each manuscript so that the cursive letters would appear in the same context.

The next step was to experiment with my data to prepare it. I used Claude to code two short Python scripts for me. The first was for binarizing the images (converting them to black and white) because I didn't have time to do it manually. The second was for resizing them and adding padding if needed.

Finally, the last step has arrived. In a notebook, I centered and converted my images to pixel vectors before calculating the cosine similarity. Since the heatmap is unreadable, I decided to use a table to display my data. I tried different ways to analyze my data to prove my hypothesis. Finally, I wondered what the ideal threshold was for defining a good score; apparently, in this case, it’s 0.75, but to be honest, I didn’t really understand it. All the code used is available in this repository.

There are limits, like the data sample is too small, using pixels as vectors isn't really ideal; it depends more on the letter's position in the image than on its shape etc. But still, the idea here was mostly to experiment.

The full detail lives in the README.