Spotlight on DIVAServices

I recently had the pleasure of visiting for the first time both the lovely island of Santorini and the Document Analysis Systems community for their DAS 2016 workshop. The visit certainly left an impression: the beauty of the scenery was complemented by a few days’ worth of interesting presentations (with slides sporting more mathematical formulae than I’ve seen in years). One of the papers in particular described a resource which I believe could yet prove useful for researchers and projects in digital humanities and editing, and which I therefore thought to share here.

DAS 2016 coffee break.

Before bringing that paper to the table, however, here’s an outsider’s sketch of a few words on the sort of work that’s done within the DAS community: their work seems to often oscillate between image processing and computer vision, relate to information retrieval and natural language processing, and centre around the algorithmic analysis of document images. In practical terms, it seems to often involve the development of methods for addressing the sorts of problems that one runs into when computing with real-life document images (e.g. document image binarisation, text line recognition), but it also concerns piecing the methods together to form larger end-to-end systems (e.g. automated document conversion). The research focus seems often not to address problems which have so far received little or no attention, but instead, to revisit important enduring problems in an effort to find improved solutions for them (e.g. nowadays many such revisits try to outperform existing solutions with deep learning-based methods). Bringing the field’s researchers together, these development efforts are often spurred on by competitions arranged on given problems and agreed datasets.

With the general sketch behind us, the particular paper I wanted to bring up here was presented by Marcel Würsch (DIVA Group, University of Fribourg). It described an online platform called DIVAServices, which is a suggestion for a more concerted way of sharing document analysis methods: the platform’s aims are to collect state-of-the-art methods, to host and share them from under a single roof, and also to facilitate a practical-level access by providing them as web services. The platform, then, aims to become a sort of living catalogue: it would be alive not only in the sense that it is to include the latest methods, but also in how the burden of setting them up-and-running would have been taken care of and they’d be executable and ready to be trialled or integrated into one’s own workflow with minimum hassle. Furthermore, the paper described a web-application called DIVAServices-Spotlight, which had been developed to showcase the platform: it is as a web-based interface offering means to survey the aggregated methods and, importantly, allowing them to be tested on one’s own document images.

While the platform is intended to serve the DAS community and to ease software engineers’ development efforts, it seems like a very welcome initiative also from a digital humanist’s perspective. From that point of view, the promise of the work lies in making both the state-of-the-art methods and their authors, hopefully also the source code and executables, much more discoverable and approachable. Indeed, it is often challenging for those who do not actively follow the field to find out what sorts of methods one can even hope to come across or to come to terms with their capabilities, let alone to determine whether to invest more effort in trying to delve deeper and set up and trial a specific method with one’s own material. I certainly had my initial troubles in getting my head around of what’s going on. In such phases of uncertainty, this platform could come in handy for newcomers: for example, they could view the available methods and apply them on a few uploaded images to quickly gain a rough idea of what kind of performance they can expect or where any specific problems lie within their particular documents and images. Such trials often prove quite revealing: for example, they can quickly squish false hopes by exposing the methods’ fragility, but also help to identify where their potential might lie. This work could therefore also end up serving as an important entry-point for people in DH to explore the DAS community’s conceptual and practical toolbox.

It’s not hard to argue for the usefulness of such an entry-point on a more general sense either, since it’d likely lead to more interaction between the two fields. During the workshop I was slightly surprised to learn that the DAS researchers mention us — people in the humanities and digital humanities — quite often as one of their methods and systems’ end-user segments. Considering that, it’s a pity that we in DH perhaps fail to acknowledge the extent to which we are already indirectly benefiting from their work (e.g. from the implementations that are tucked away within the processing pipelines of larger systems, such as in library digitisation workflows). It’s also becoming evident that many in DH are nowadays sitting on piles and piles of document images and quite naturally growing more and more interested in making a fuller use of them; were the methods catering for such purposes slightly more approachable, many might jump at the opportunity for finding uses for them. In the end, one cannot help but to feel that good things might one day come from an invested collaboration between those who have a research interest and expertise in documents, and those who have the same in interfacing them through their digital images.

The work on DIVAServices seems still to be in its early phases. Nevertheless, now is probably a good time to voice any interest towards it: these sorts of initiatives seem often to be at a risk of running out of steam and into a quiet neglect all too soon after the early interest wanes. In case you’re eager to find out more, make sure you catch Marcel’s talk at DH2016 in Kraków (the preliminary schedule places it in the morning session of Wed 13th July). Notably, Marcel’s presentation at DAS 2016 won him the Best Student Paper Award.

—

Würsch, Marcel, Rolf Ingold, and Marcus Liwicki. “SDK Reinvented: Document Image Analysis Methods as RESTful Web Services”. 90–95. 12th IAPR Workshop on Document Analysis Systems, Santorini, 2016.

Würsch, Marcel, Michael Bärtschi, Rolf Ingold, and Marcus Liwicki. “DIVAServices-Spotlight – Experimenting with Document Image Analysis Methods in the Web”. Digital Humanities 2016, Kraków.