Sharing Data for Handwritten Text Recognition (HTR) - Archive ouverte HAL Accéder directement au contenu
Chapitre D'ouvrage Année : 2024

Sharing Data for Handwritten Text Recognition (HTR)

Résumé

Handwritten Text Recognition (HTR) is at present perhaps the principal application of Artificial Intelligence to the Digital Humanities. It falls under the category of supervised machine learning, and this in turn depends almost entirely on the data that is used for training. The consequences of this are numerous: these techniques are therefore positivist, insofar as they are only applicable to cases where the answer can be known and defined in advance; they will necessarily reflect historical practice; and they will also reflect all of the inevitable biases that are present in the data (for just one example of which see Brown et al., 2020, 36-39). The data itself also becomes valuable and so a commodity in its own right and, conversely, the availability or lack of data is itself shaping decisions about applications in machine learning. For HTR, this means that the most progress has been on modern material in widelyused languages and writing systems such as English and others written in the Latin alphabet, while socalled rare and historical scripts have seen much less success. It also suggests many communal benefits in publishing and sharing training data, in order to combine effort and expertise and avoid unhelpful repetition of labor. Such sharing brings many challenges, as it requires standards or at least common practices in transcription (including treatment of abbreviations, punctuation, "non-standard" spelling and capitalization and so on), as well as standards for data sharing that allow for the many important variations in the world's writing systems. It also assumes the willingness (and the possibility) of sharing data openly, including transcriptions and images, but this in turn can depend on many different institutions and interests. Despite these challenges, some initiatives have begun to point the way, and the benefits of this work are already being felt, which is encouraging for the future of the field.
Fichier principal
Vignette du fichier
dh_in_practice_preprint.pdf (105.45 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04444641 , version 1 (07-02-2024)

Identifiants

  • HAL Id : hal-04444641 , version 1

Citer

Peter Stokes, Benjamin Kiessling. Sharing Data for Handwritten Text Recognition (HTR). Digital Humanities in Practice, In press. ⟨hal-04444641⟩
11 Consultations
30 Téléchargements

Partager

Gmail Facebook X LinkedIn More