Assignment 2: PG 93 col. 1571-2

In this forum, we will work with people who are training OCR engines for Polytonic Greek, helping them improve the quality of their engines by correcting specific texts.
Jonathan Robie
Posts: 3486
Joined: May 5th, 2011, 5:34 pm
Location: Durham, NC
Contact:

Re: Assignment 2: PG 93 col. 1571-2

Post by Jonathan Robie » May 24th, 2014, 9:05 am

BruceRobertson wrote:Well, the old joke is that if you create a number of pages of ground truth = to the number of pages in the target corpus, you should get pretty close to perfect results :-) However, if my calculations are correct, the existing classifier is trained on 8 whole pages of ground truth. So any additional pages, especially those that contain lots of characters not well represented in those 8 pages, will help quite a bit.
Roughly how many pages of Ground Truth are you likely to want for Migne? I'm trying to scope the work.
0 x


ἐξίσταντο δὲ πάντες καὶ διηποροῦντο, ἄλλος πρὸς ἄλλον λέγοντες, τί θέλει τοῦτο εἶναι;
http://jonathanrobie.biblicalhumanities.org/

BruceRobertson
Posts: 22
Joined: May 16th, 2014, 9:49 pm

Re: Assignment 2: PG 93 col. 1571-2

Post by BruceRobertson » May 24th, 2014, 1:46 pm

Here's a sort of rule-of-thumb: assuming we'll never get up to 100%, each doubling of ground truth gets us twice as close to, say 98% character accuracy. So if we're at 93% now, ten more pages might get us to 95%. I think ultimately, given how large and diverse Migne is, I'd like to have about 50 pages of gt. But these would have to be from well chosen pages (random sample).

The right way to go about this is to provide a substantially larger amount of gt, retrain with that, and then use the resulting output from that to more easily provide even more gt, etc. I won't retrain and re-process with just two more pages of gt, but I will after we get, say 8 or more, because that will make a difference.
0 x

Post Reply