Irish monks in the sixth century A.D. who were unfamiliar with the Latin language from the continent first introduced one of the most important tags still in use today – spaces between words*. Standardized punctuation quickly followed[*]. Later, during the Middle Ages, a kind of PML (Paragraph Markup Language) was introduced, similar to today’s SGML but back then only opening tags indicated the beginning of a paragraph. It was not until the 15th century that document layout became what we are familiar with today.
So while we think of markup as modern, it originally was created to help humans read, and has since evolved into an essential tool for computer processing. Where this evolution is headed is important, given the growth of documents and their relationship to automated workflows. While the production of documents is not growing at quite the same exponential pace of numerical data, it is still huge and critical to many businesses. Just as the monks relied on tags to help humans read, the effort to “teach” computers how to structure documents will lead to more efficient workflows.
Markup has a long tradition in print shops where it is used to format and correct manuscripts. This tradition was passed over into photocomposition before computation with languages such as:
<CC 0,5, 12>Nortext,
:li.Ordered GML (Generalized Markup Language)
.bu .b Troff[†].
These first computed markup languages were more or less a copy of existing manual practices, and corresponded with textual streams where procedural tags about layout were added to explain how content must be laid out. At the end of the 1960s, under the impetus of the Graphic Communication Association the GenCode (Generic Coding) Committee was created with the goal of reflecting on the separation of document information content from document format. Their conclusions, embodied in the "GenCode® concept", were that:
Furthermore, markup languages [should] greatly facilitate the sharing of data and the integration of diverse types of software, yielding a new era of efficiency and flexibility.”[‡]
The first technological achievement that put these principles into practice was GML (Generalized Markup Language) designed by C. F. Goldfarb, E. Mosher and R. Lorie in the 1970s. Then followed its offspring: SGML, HTML and XML. XML, which leveraged the experience of SGML but which was easier to manipulate, emerged as an industry standard and is recognised as “the universal format for structured documents and data on the Web.”[§] Well specified and based on international standards such as Unicode, a set of tools was quickly developed by the community and made available to users. In line with the first GenCode concept of incorporating smaller documents within bigger ones, specific languages were also designed including SMIL for multimedia data, RDF for resources or MathML for mathematical formulae.
Following the second GenCode concept, a structured document is one that is hierarchically organized by elements explicitly marked up with a tag. This markup, which reflects the semantic structure of the document, does not correspond to layout instructions. The layout instructions are given by a style sheet (XSL for instance), which indicates how a tag (e.g., its content) must be laid out.
Finally, a document type called schema, may be associated with a document, so that it can be validated e.g., to check it complies with a specific model. This warranty prevents inconsistency in further processing. This separation of content from form is a benefit for publishers who now target multiple output channels. It is also essential in automated document processing to:
It would be reasonable to assume that anyone who is literate can write a digital document using simple word processing software. But who is able to actually design a markup document? To do so one must select the correct document model, properly markup its pieces, validate compliance against its chosen model and select the formatting style sheet to display it. After having worked 20 years on markup, Brian Reid, father of the Scribe system, presented his reflections on markup technologies in 1998 one of which is still valid to this day: Most people won’t use abstract markup even if you threaten them.[**] Document markup is beyond a layman’s skills and is reserved for professionals[††]. Word processing software provides some assistance with spell checkers and grammar checkers. Yet today’s markup checkers are at the level of the very first spell checkers from the 1960s which could only look up a word to see if it existed in a checklist. Unstructured documents accumulate in our PCs every day and the “new era of efficiency and flexibility” that markup could bring is still a dream.
If most of these documents are unstructured, how can we automate their processing? How can they be easily structured to be fed into automated workflows? If humans are not ready to accept this burden, computers have to be taught to undertake it.Over the years, a team at XRCE has been investigating the challenges related to the automation of document processes, including document understanding, document conversion to XML and schema management. The team addresses research challenges relevant to analyzing and understanding document collections based on their layout and structural organization. These methods can be thought of as going beyond what Optical Character Recognition systems do at the character and word level to reconstruct higher level structures, such as document sections, tables of contents, indexes, etc. to automatically markup documents.
The main difficulty lies in the enormous variety of document content layout. To exhaustively inventory all the possibilities is a never ending and expensive task. One alternative is to inverse this pattern recognition problem by relying on constraints that structure the different elements in a given document (for instance the incremental relation between two page numbers), without having to describe their layout characteristics. [‡‡] We simply learn for each document which layout has been used, and mark it up accordingly. These methods have been successfully applied to large-scale customer cases where required information has been extracted from millions of pages, automatically feeding customer databases. Some examples of this technology are available as online web services on Open Xerox including the popular Pdf2epub converter . This service will automatically convert for you your PDF file into an ePub file, so that reading will be optimal on tablets and readers.
About the author:
Hervé Déjean is a senior scientist at Xerox Research Centre Europe. His research interests lie in structural pattern recognition and his primary objects of study are texts and documents. He is one of the inventors of the Xerox Pdf2epub converter.