Project Update – Bentham vs the computer
By Louise Seaward, on 23 February 2018
Throughout it’s long history, the Bentham Project has always been interested in the way in which technological advances could be integrated into its work on the scholarly edition of Bentham’s Collected Works. Transcribe Bentham is currently a proud partner in an international collaboration focused on using innovative computer science techniques to process historical manuscripts. The mission of the READ (Recognition and Enrichment of Archival Documents) project is to make archival collections more accessible through the development and dissemination of Handwritten Text Recognition (HTR) technology.
This technology is freely available through the Transkribus platform. Using algorithms of machine learning, it is possible to teach a computer to read a particular kind of handwriting – even Bentham’s! The technology is trained by being shown images of documents and their accurate transcriptions. Thanks to the hard work of the Transcribe Bentham volunteers, we are lucky to have a sizeable collection of transcripts that can be used as training data for automated text recognition.
We last summarised our experiments with HTR in a blog post from June 2017. At that point, we had used technology from the Computational Intelligence Technology Lab (CITlab) at the University of Rostock to produce a model capable of processing the easier papers from the Bentham collection, largely those written by Bentham’s secretaries. This model can automatically produce transcripts with a Character Error Rate of between 5 and 10%, meaning that 90-95% of characters in the transcript are correct. The Bentham model is now publicly available in Transkribus under the title ‘English Writing M1’ and has been applied to other collections of eighteenth- and nineteenth-century English handwriting with some success.
Although this model copes well with documents where the handwriting and layout are relatively clear, it struggles to recognise the more difficult examples of writing from Bentham’s own hand. So we decided to take on the challenge of teaching a computer to read some of the very worst examples of Bentham’s handwriting!
We used the Transkribus platform to create training data based on Boxes 30 and 31 of the Bentham Papers held in UCL Special Collections. These manuscripts were part of Bentham’s lifelong obsession with critiquing the work of William Blackstone, the English jurist who was most famous for his Commentaries on the Laws of England (1765-9). To create the training data, we uploaded around 200 digital images to Transkribus, segmented each image into lines and then copied over existing transcripts to match each image.
The resulting model generates transcripts with an average Character Error Rate of 26%. This error rate is unfortunately too high to automatically produce transcripts suitable for scholarly editing. Nevertheless, it does have the potential to facilitate the full-text search of the Bentham Papers. Transkribus now includes sophisticated Keyword Spotting technology, which is capable of finding words and phrases in documents, even if they have been mistranscribed by the computer.
We are working with the Pattern Recognition and Human Language Technology (PRHLT) research centre at the Universitat Politècnica de València and the Digitisation and Digital Preservation group at the University of Innsbruck to present an open-access search functionality for the Bentham Papers. We are also hoping that volunteers could get involved in this endeavour by checking and correcting the results of significant search queries to ensure their accuracy.
Improving the recognition of Bentham’s handwriting is our other aim and to this end, we will be producing more pages of training data in Transkribus. The technology moves so fast that the efficiency of this process has already been streamlined thanks to technology from CITlab. Transkribus can now easily find lines in images (even in documents with complex layouts) and it is also possible to use existing transcripts to automatically train a model, rather than copying them into Transkribus line by line.
I would like to thank Chris Riley, PhD student and transcription assistant at the Bentham Project, for helping to produce the training data for the latest Bentham model.