Joho the Blog » [2b2k] Crowdsourcing transcription
Everyday Chaos
Too Big to Know
Too Big to Know
Cluetrain 10th Anniversary edition
Cluetrain 10th Anniversary
Everything Is Miscellaneous
Everything Is Miscellaneous
Small Pieces cover
Small Pieces Loosely Joined
Cluetrain cover
Cluetrain Manifesto
My face
Speaker info
Who am I? (Blog Disclosure Form) Copy this link as RSS address Atom Feed

[2b2k] Crowdsourcing transcription

[This article is also posted at Digital Scholarship@Harvard.]

Marc Parry has an excellent article at the Chronicle of Higher Ed about using crowdsourcing to make archives more digitally useful:

Many people have taken part in crowdsourced science research, volunteering to classify galaxies, fold proteins, or transcribe old weather information from wartime ship logs for use in climate modeling. These days humanists are increasingly throwing open the digital gates, too. Civil War-era diaries, historical menus, the papers of the English philosopher Jeremy Bentham—all have been made available to volunteer transcribers in recent years. In January the National Archives released its own cache of documents to the crowd via its Citizen Archivist Dashboard, a collection that includes letters to a Civil War spy, suffrage petitions, and fugitive-slave case files.

Marc cites an article [full text] in Literary & Linguistic Computing that found that team members could have completed the transcription of works by Jeremy Bentham faster if they had devoted themselves to that task instead of managing the crowd of volunteer transcribers. Here are some more details about the project and its negative finding, based on the article in L&LC.

The project was supported by a grant of £262,673 from the Arts and Humanities Research Council, for 12 months, which included the cost of digitizing the material and creating the transcription tools. The end result was text marked up with TEI-compliant XML that can be easily interpreted and rendered by other apps.

During a six-month period, 1,207 volunteers registered, who together transcribed 1,009 manuscripts. 21% of those registered users actually did some transcribing. 2.7% of the transcribers produced 70% of all the transcribed manuscripts. (These numbers refer to the period before the New York Times publicized the project.)

Of the manuscripts transcribed, 56% were “deemed to be complete.” But the team was quite happy with the progress the volunteers made:

Over the testing period as a whole, volunteers transcribed an average of thirty-five manuscripts each week; if this rate were to be maintained, then 1,820 transcripts would be produced every twelve months. Taking Bentham’s difficult handwriting, the complexity and length of the manuscripts, and the text-encoding into consideration, the volume of work carried out by Transcribe Bentham volunteers is quite remarkable

Still, as Marc points out, two Research Associates spent considerable time moderating the volunteers and providing the quality control required before certifying a document as done. The L&LC article estimates that RA’s could have transcribed 400 transcripts per month, 2.5x faster than the pace of the volunteers. But, the volunteers got better as they were more experienced, and improvements to the transcription software might make quality control less of an issue.

The L&LC article suggests two additional reasons why the project might be considered a success. First, it generated lots of publicity about the Bentham collection. Second, “no funding body would ever provide a grant for mere transcription alone.” But both of these reasons depend upon crowdsourcing being a novelty. At some point, it will not be.

Based on the Bentham project’s experience, it seems to me there are a few plausible possibilities for crowdsourcing transcription to become practical: First, as the article notes, if the project had continued, the volunteers might have gotten substantially more productive and more accurate. Second, better software might drive down the need for extensive moderation, as the article suggests. Third, there may be a better way to structure the crowd’s participation. For example, it might be practical to use Amazon Mechanical Turk to pay the crowd to do two or three independent passes over the content, which can then be compared for accuracy. Fourth, algorithmic transcription might get good enough that there’s less for humans to do. Fifth, someone might invent something incredibly clever that increases the accuracy of the crowdsourced transcriptions. In fact, someone already has: reCAPTCHA transcribes tens of millions of words every day. So you never know what our clever species will come up with.

For now, though, the results of the Bentham project cannot be encouraging for those looking for a pragmatic way to generate high-quality transcriptions rapidly.

Previous: « || Next: »

Leave a Reply

Comments (RSS).  RSS icon