Over the last year or so, the IHR Digital History seminar has become increasingly web-focussed, which is of course of interest to me (if not necessarily to everyone.) Last week we had an excellent paper from Jack Grieve of Aston University on the tracking of newly emerging words as they appeared in large corpora of tweets from the UK and the US. By amassing very large tweet datasets, he and his colleagues are able to observe the early traces of newly emerging words, and also (when those tweets were submitted from devices which attach geo-references) to see where those new words first appear, and how they spread. Jack and his colleagues are finding that words quite often emerge first (in the US) in the east and south-east (or California) and then spread towards the centre of the continent. They don’t necessarily spread in even waves across space, or even spring between urban centres and then to rural areas (as would have been my uneducated guess). Read more at the project site, treets.net, or watch the paper.
This kind of approach is quite impossible without the kind of very large-scale natural language data such as social media afford. This is particularly so as most words are (perhaps counter-intuitively) rather rare. In the corpus in question, the majority of the 67,000 most common words appear only once in 25 million words. Given this, datasets of billions of tweets are the minimum size necessary to be able to see the patterns.
It was interesting to me as a convenor to see the rather different spread of people who came to this paper, as opposed to the more usual digital history work the seminar showcases. Jack focussed on tweets posted since 2013; a time span that even the most contemporary historian would struggle to call their own; and so not so many of them came along – but we had perhaps our first mathematician instead. This was a shame, as Jack’s paper was a fascinating glimpse into the way that historical linguistics, and indeed other types of historical enquiry, might look in a couple of decades’ time.
But there is a caveat to this, which was beyond the scope of Jack’s paper, to do with the means by which this data will be accessible to scholars of 2014 working in (say) 2044. Jack and his colleagues work directly from the so-called Twitter “firehose”; they harvest every tweet coming from the Twitter API, and (on their own hardware) process each tweet and discard those that are not geo-coded to within the study area. This kind of work involves considerable local computing firepower, and (more importantly) is concerned with the now. It creates data in real time to answer questions of the very recent past.
Researchers working in 2044 and interested in 2014 may well be able to re-use this particular bespoke dataset (assuming it is preserved – a different matter of research data management, for another post sometime). However, they may equally well want to ask completely different questions, and so need data prepared in a quite different way. Right now, the future of the vast ocean of past tweets is not certain; and so it is not clear whether the scholar of 2044 will be able to create their own bespoke subset of data from the archive. The Library of Congress, to be sure, are receiving an archive of data from Twitter; but the access arrangements for this data are not clear, and (at present) are zero. So, in the same way that historians need to take some ownership of the future of the archived web, we need to become much more concerned about the future of social media: the primary sources that our graduate students, and their graduate students in turn, will need to work with two generations down the line.
Certainly, historians have always been used to working around and across the gaps in the historical record; it’s part of the basic skillset, to deal with the fragmentary survival of the record. But there is right now a moment in which major strategic decisions are to be made about that survival, and historians need to make themselves heard.
[This post also appears on the IHR Digital History Seminar blog.]