Telegraph shown


Already mapped lines

#where are stations? where do lines really intersect and where are transfers harder
#but who operates?
#where are these on a yearly basis? 
# consolidation of railroad ownership

One way to answer:

Official Guides to the Railway

# sometimes monthly, quarterly trade publication published by # variety of information on railroads, steam ships, schedules, fares, maps. # images # index of railway stations: lists places, companies operating there, # images # list of companies: old and new owner/operating companies # images

What we can learn:

# for each year (and often multiple times per year) # station location # companies serving that station # are stations in a place nearby or how far apart (if they do not intersect) # what previous roads are now operated/owned by other companies


Other methods:

Direct transcription by team of RAs

# somewhat costly, harder to ensure standardization, possibility of transcription errors, wrong lines, etc # esp. given: images are large, lots of lines of text, hard to read, easy to lose track


# OCR works to some extent, but # small text, poor printing, #even with training on in-document characters training, performance requires massive cleaning # can do this with RAs again, but hard to get high quality… not all errors automatically detected, hard to parse (esp. with commas, periods, etc)


# problems similar across railroad and telegraph images # how does zooniverse provide a solution?

# 1. more workers, volunteers rather than hired help. costs are less in the long run # 2. easy process for having multiple transcriptions of the same item, improving quality # 3. flexible set of tools, possible to customize rules for allocating images to tasks for completion # quite a different project than the others we’ll see # but makes use of entry forms for each image, much easier to produce this here than on your own with google forms, amazon turk # 4. Discretized tasks: division of labor across individuals, or easier to divide labor across time (helpful even if we use RAs on this service)


Key customization:

# needed some way to get subset of image # possible to have volunteers mark columns, rows, then use image software to split # too many rows

# need some automation # automatic splitting into columns performed poorly: not always straight columns, sometimes splitting wrong places, not reliable # but, within columns, splitting into rows automatically performs well

# users identify what kind of page we have/state in case of telegarph # repeated classification of these splits, cluster these splits and use that to split images into columns # then can use image analysis to split column into rows # results in a list of all cells on a page, retaining column index and row index on the column. # new cell images, retaining page and page position metadata are then passed to back zooniverse for classification/transcription

# this permits us to shows workers one small image, reduce complexity of the task given to them # downside: need to write guides so people can understand the information even when it lacks full context # downside: need to pull information back together, but this uses row index and questoins we ask

this was the biggest challenge:

# could not do it internally to zooniverse, and ensuring that this process works without problems required substantial coding help # now that it is completed, any other documents with similar tabular structure can be divided up

Subsequent workflows

# Railroad row classification # what is on this? can then prioritize information we transcribe # Telegraph row classification #

Examples of current workflows

How to make people participate


data completion:

Longer term:

Advantage of crowdsourcing is you attract people to

Long term