By Brian Guo ’25
I worked with two relatively modern, handwritten texts that are connected to the Brut (I believe they came together with this manuscript) in order to test out Transkribus and just provide/decipher textual data of these texts. Different from if I would’ve wanted to test out Transkribus’s effectiveness on medieval texts, Transkribus has a few public models that are suitable for modern English, so that I would be able to test out how these public models work. These 20th century scribbled handwriting would post a decent challenge to a tool that focuses on getting machines to read texts. These are also very short texts – two pages each – so that training a private model that requires a fair amount of ground truths wouldn’t be suitable. I’m also interested in how these already trained models could be applied for a text that is more or less independent – texts that are not directly related to the material that a model has been trained with so that it isn’t “ready” for it. I want to see how the tool can help me start working with a text that is less difficult to work with than the original Brut.
Transkribus, model: The Text Titan I, which is applicable for German, Dutch; Flemish, French, Finnish, Swedish and English
For the purpose of these two texts I am using a few public models with relatively low CER (character error rate). Genuinely, the performance of these models aren’t that different in terms of level. The English Eagle seems to have a higher CER at 8%, but it generally does a similar job with the ones that are around 2%. None of the models are accurate enough to depend upon, as is expected.
It is evident that Transkribus performs a much better job with a modern text. We can see that Transkribus is already giving some remotely fluent/eloquent sentences, like “Brut Chronide- English Ms on vellum, # 2 5 + 116 leaves (293 by 203 mm), headings of # 3 chosters in red, initials in red and blec, # 4 with pen work ornamentation carried into the # 5 margin 15th century.”. It sometimes proves exceptionally useful at providing me some sort of context to interpret the language through a different lens. For instance, I couldn’t make it out that the letters after “293 by 203” are “mm”, because I was trying to interpret it as a word rather than some symbols or abbreviations.
I shared my experience with Professor Warren, who told me that transcription might initially sound like looking at the text word by word – or rather letter by letter – and figuring out what each letter is. In practice, however, transcription works more like a mix of a series of recognition and guessing, in which the transcriber really needs to read the text itself and read outside the text.
Transkribus encountered some major difficulties when the handwriting gets more scribbled in the latter half of the text:
So I’ll have to look at this part more by myself and follow some of the clues that will lead me to recognizing the text. For instance, the year number 1419 has led me to recognizing that the writer is referring to the siege of Rouen.
The text also refers to some medieval slangs found in the marginal inscriptions/annotations in the book which I cannot possibly decipher by myself. I was able, however, to find information regarding the Dartmouth Brut on the website of an auction house, Christie’s. The web link is pasted below. It notes these inscriptions like “in paule mier is, in myer eill is, in oke tymber is, in ferne non is”. I’m not sure whether this online description is referring to the text I’m transcribing or rather the other reader was just reading the manuscript itself. Having this online source of reference does help me a lot.
The URL of the Christies site:
https://www.christies.com/en/lot/lot-1837718?lid=4&sc_lang=zh-cn
The other text was a letter written by S. Gibson in 1922. The later descriptive note also refers to the Gibson letter. Having the Gibson letter available also helps me decipher the later note.
Transcribing both texts took me a few hours. The process involves testing out different public models on Transkribus to piece together different missing parts, looking carefully at the machine-generated text to correct them, and searching online + comparing the two texts to each other to read around the text.
Currently, I have a transcribed version of each secondary text, but they are by no means finished. There are still details (one or two per text) that I couldn’t be sure of. The transcription has also not been reviewed by another reader, but I consider it mostly done. I’m going to paste them here:
Text 1:
Brut Chronicle – English MS in English on vellum, 5+116 leaves (293 by 203 mm), headings of chapters in red, initials in red and blue, with pen-work ornamentation carried into the margin. 15th century.
This famous English Chronicle, called the “Brut” or “Dunstable” Chronicle, contain extends from of the live of the legendary Brute (Brutus?) to Henry V. The first 5 leaves of the present MS contain a table of chapters in a hand slightly later than the text. The first leaf of the text bears the remains of a fire illuminated border, but the upper part of the leaf is defective, and it has shrunk: a portion of one leaf in gathering 6 has been torn away. and the 4 middle leaves of the last gathering are wanting: one gathering is apparently wanting at the end. The last leaf extant relates the preparation for the siege of Rouen (1419). Only one leaf of text is wanting at the end, as this Recension ends with the siege of Rouen there was
probably also a blank leaf at the end wanting.
There are several notes of early ownership: the earliest seems to be Thomas Wykyng of Cowden (no doubt the village in Kent): the next takes the MS to Durham William Ewre of Bushopp Midleham Esqr oweth this Booke. T(i.e. teste) Clement Barnes. Other notes are “Richard Smyth of Quodington”, “Amen quod Wyllmd Harper”, “Rynnard Russel”.
There are several early inscriptions in the margins, e.g. the pathetic complaint “It ys to harde for my lernyng”; others seem to be popular sayings “In paule mier is, in myer eill is, in oke tymber is, in ferne non is”.
The MS is bound in tight/light leather binding not strengthened in any way and with a folding flop which has been repaired; the Roll used on the binding varies it is a very early one and I am unable to name it. This form of unstrengthened binding is very uncommon and in this case I should say dated from about 1450 to 1500 probably about the date the table was written.
but S. Gibson dates the binding C15-50 xx
Text 2
Dear Mr. Bosanquet,
Please accept my best thanks for sending those two very interesting bindings for me to see.
The binding of the Brut is of a most uncommon type. I do not recognize it as Oxford work. I should say that it is English work of c. 1550. It cannot possibly have been executed before 1525. The text of the MS seems to be of the 1st quarter of the 15th century and the index of about 50 years later.
I recognized the binding of the other book at once a Cambridge work. The use of red enamel or paint on Cambridge bindings is very common at a certain period. It is then found on a class of bindings which has a bird like a pheasant as its distinction feature. It is therefore unsafe to assume that the introduction of color is due to feminine Làrlè. However whether Little Gidding or “Buck” it’s a charming little book.`
The books return to-day by registered post.
Yours Sincerely
S. Gibson
