Written language is segmented in various ways. White space separates one word from another. Non-alphabetical symbols identify clauses (commas) and sentences (periods). Question or exclamation marks identify certain kinds of sentences. White space of different kinds (indentation, blank lines) is used to articulate paragraphs as the fundamental discursive units. Hierarchically ordered larger units with special headings identify sections and chapters. Familiarity with this system of ordering devices is part of basic reading competence.
Because humans are slow but smart, they need to be reminded only once of these points of discursive articulation. The same comma or period ends one bit of language and starts another. In Spanish, questions and explanation are explicitly marked both at their beginning and their end with upended and normal versions of the punctuation mark: ¿How are you? ¡Don’t ask! You can think of the opening and closing punctuation defining a two-dimensional container with the question or exclamation as its content.
This is exactly what happens in text encoding, where a text is defined as an “ordered hierarchy of content objects” (OHCO) and a system of hierarchically ordered opening and closing boundary markers creates containers whose content can be identified by a machine and processed in various ways. “Element” is the technical term for such containers. By convention the angle brackets familiar from HTML have become the universal boundary markers. In principle, it could be anything. The important thing to remember is that if a stretch of text is unambiguously defined in terms of its beginning and end, you can tell a machine to identify it and do this or that with it. While today’s search engines still do not provide much support for “element aware” searching, the coarse but quite consistent encoding that the TCP has implemented across more than 50,000 texts will in the years to come be an increasingly valuable feature of this corpus. For instance, it makes it very easy to look for words that occur in poetry or for words that only occur in poetry.
The EarlyPrint versions add an additional encoding layer. The entire TCP corpus has been annoted with MorphAdorner, a program developed by Phil Burns. In this procedure every word is wrapped in a <w> element and given “attributes” that specify its part of speech, lemma or dictionary entry form, and a standardized form of its spelling. Thus a spelling like “louyth” would be encoded as
<w lemma=”love” pos=”vvz” reg=”loveth”>louyth </w>
Human readers never get to see this ugly “explicitation” that tells them only what they know already or would immediately gather from the reading context. But it enables a search engine to retrieve ‘louyth’ in a search for all forms of ‘love’. Tagging of this kind increases the size of the file by an order of magnitude and makes its raw form almost unreadable for humans. But it allows a machine to see through the surface forms of words and recognize more abstract lexical or grammatical patterns. Linguistic annotation allows you to identify patterns that are defined syntactically rather than lexically. The opening sentence of Jane Austen’s Emma includes a famous example of the three-adjective rule:
Emma Woodhouse, handsome, clever, and rich.
A search for that pattern in Early Modern plays retrieves lines like
The Scottish king grows dull, frosty, and wayward.
Given the high degree of orthographic variance in the Early Modern world, linguistic tagging that is consistent and stays below a tolerable error rate of ~3% is a critical component in achieving “agile data integration” in the EarlyPrint corpus.
The combination of linguistic annotation with the existing encoding allows for clearer articulation of some textual features. There are on the order of ten million Latin words in the TCP texts. They are typically printed in italics and therefore wrapped in <hi> tags by the transcribers. If the tokens inside <hi> tags consist mainly of words that the linguistic annotation identified as Latin, you can retag the stretch with the semantically more epressive tag <foreign xml:lang=”lat”>. In a similar procedure you add a language attribute to elements where a high percentage of the words are tagged as Latin. We have done this, and the error rate is very low. Isolating non-English content is useful even if you are not interested in it, because it defines the English content more clearly. Because there is so much Latin in the TCP corpus, it is easy to extract with high accuracy. Other languages are more problematical. Short passages of French are particular difficult because Early Modern English and French orthography overlap in many ways.
Linguistic annotation is also of great help in the identification of “named entities”, which include not only single names but also word strings the components of which are not names (United Provinces).The Weimar edition of the works of Luther runs to 60 volumes plus a dozen index volumes for people, places, things, and citations. Linguistic annotation, combined with the existing encoding, will make it possible to create corpus-wide indexes at a fraction of the time it took to create the index volumes to the Luther edition. These digital indexes will not be as delicately crafted, but they will cover a lot more ground, and there will be few purposes for which they do not offer significant help.