Print

Print


Awesome find!

Mike

From: Geospatial Hosting Environment <[log in to unmask]> On Behalf Of Hessler, John W
Sent: Friday, May 04, 2018 7:51 AM
To: [log in to unmask]
Subject: Toponym matching...using DNN

GHErs,

During yesterday's sessions on machine and deep learning, Mike Schoelen brought up the idea of name and place recognition...a sort of OCR for maps....I thought I would send to all of you an interesting paper from the International Journal of Geographical Information Science, from a few months ago, where they are using deep neural nets...in this case recurrent networks, which are more useful for language strings, as opposed to the convolutional nets you saw yesterday that tend to be for images...there has been a great deal of progress on this in the last few years and in fact the team from Voyager GIS has had some success in string comparison...there are a few groups like Google who are using bi-directional encoders and quantized training and have really improved machine translations and which also has applicability here...

I think at the LoC the auto-metadata generating capabilities of recurrent nets would be more the direction for our historical materials and increase searchability...

It is a nice article with a good bibliography....happy reading....


John

John Hessler, FRGS
Specialist in Mathematical Cartography &
Geographic Information Science

Geography and Map Division
Library of Congress
Washington, DC
202-707-7223
[log in to unmask]<mailto:[log in to unmask]>

Twitter: https://twitter.com/topology_lab
Center for Open Science Page: https://osf.io/zt8c4