Digital scholarship blog

20 January 2020

Using Transkribus for Arabic Handwritten Text Recognition

This blog post is by Dr Adi Keinan-Schoonbaert, Digital Curator for Asian and African Collections, British Library. She's on Twitter as @BL_AdiKS.

 

In the last couple of years we’ve teamed up with PRImA Research Lab in Salford to run competitions for automating the transcription of Arabic manuscripts (RASM2018 and RASM2019), in an ongoing effort to identify good solutions for Arabic Handwritten Text Recognition (HTR).

I’ve been curious to test our Arabic materials with Transkribus – one of the leading tools for automating the recognition of historical documents. We’ve already tried it out on items from the Library’s India Office collection as well as early Bengali printed books, and we were pleased with the results. Several months ago the British Library joined the READ-COOP – the cooperative taking up the development of Transkribus – as a founding member.

As with other HTR tools, Transkribus’ HTR+ engine cannot start automatic transcription straight away, but first needs to be trained on a specific type of script and handwriting. This is achieved by creating a training dataset – a transcription of the text on each page, as accurate as possible, and a segmentation of the page into text areas and line, demarcating the exact location of the text. Training sets are therefore comprised of a set of images and an equivalent set of XML files, containing the location and transcription of the text.

A screenshot from Transkribus, showing the segmentation and transcription of a page from Add MS 7474
A screenshot from Transkribus, showing the segmentation and transcription of a page from Add MS 7474.

 

This process can be done in Transkribus, but in this case I already had a training set created using PRImA’s software Aletheia. I used the dataset created for the competitions mentioned above: 120 transcribed and ground-truthed pages from eight manuscripts digitised and made available through QDL. This dataset is now freely accessible through the British Library’s Research Repository.

Transkribus recommends creating a training set of at least 75 pages (between 5,000 and 15,000 words), however I was interested to find out a few things. First, the methods submitted for the RASM2019 competition worked on a training set of 20 pages, with an evaluation set of 100 pages. Therefore, I wanted to see how Transkribus’ HTR+ engine dealt with the same scenario. It should be noted that the RASM2019 methods were evaluated using PRImA’s evaluation methods, and this is not the case with Transkribus evaluation method – therefore, the results shown here are not accurately comparable, but give some idea on how Transkribus performed on the same training set.

I created four different models to see how Transkribus’ recognition algorithms deal with a growing training set. The models were created as follows:

  • Training model of 20 pages, and evaluation set of 100 pages
  • Training model of 50 pages, and evaluation set of 70 pages
  • Training model of 75 pages, and evaluation set of 45 pages
  • Training model of 100 pages, and evaluation set of 20 pages

The graphs below show each of the four iterations, from top to bottom:

CER of 26.80% for a training set of 20 pages

CER of 19.27% for a training set of 50 pages

CER of 15.10% for a training set of 75 pages

CER of 13.57% for a training set of 100 pages

The results can be summed up in a table:

Training Set (pp.)

Evaluation Set (pp.)

Character Error Rate (CER)

Character Accuracy

20

100

26.80%

73.20%

50

70

19.27%

80.73%

75

45

15.10%

84.9%

100

20

13.57%

86.43%

 

Indeed the accuracy improved with each iteration of training – the more training data the neural networks in Transkribus’ HTR+ engine have, the better the results. With a training set of a 100 pages, Transkribus managed to automatically transcribe the rest of the 20 pages with 86.43% accuracy rate – which is pretty good for historical handwritten Arabic script.

As a next step, we could consider (1) adding more ground-truthed pages from our manuscripts to increase the size of the training set, and by that improve HTR accuracy; (2) adding other open ground truth datasets of handwritten Arabic to the existing training set, and checking whether this improves HTR accuracy; and (3) running a few manuscripts from QDL through Transkribus to see how its HTR+ engine transcribes them. If accuracy is satisfactory, we could see how to scale this up and make those transcriptions openly available and easily accessible.

In the meantime, I’m looking forward to participating at the OpenITI AOCP workshop entitled “OCR and Digital Text Production: Learning from the Past, Fostering Collaboration and Coordination for the Future,” taking place at the University of Maryland next week, and catching up with colleagues on all things Arabic OCR/HTR!

 

Digital scholarship blog recent posts

Archives

Tags

Other British Library blogs