I'm not sure but in theory I would expect the layouts to be less variable with magazines due to the smaller page space. And vLLMs seem to be good at odd text inside images
I think I see the confusing part now. The /subscribe page only is visible once you google auth or enter email / confirm your email exists. It's $9.99/month for the default selection though. In stripe it has the no CC option, so if you click the button it will just auto cancel after a week anyway
I'm going to try to figure out how to make a subset of the data searchable for free (for anyone anonymous or not without having to start free trial) as well in the next few days, so people can play with that too before committing to anything.
When you're anonymous or google/email authd, you can now see today-in-history stuff. There should be a gentle nudge to sign up when you view a component or go to pages that require activating the free trial (no cc required) which opens up search / sleuth / custom collections. That takes you to /subscribe with all the options
I'm opening up https://snewpapers.com/today-in-history to the world right now. Per @benwillis's advice below, I will figure out how to make a section of the data searchable for free as well, but this is the best I can do today!
Thank you everyone for taking the time to look at snewpapers today, enjoy!
I see what you mean now with the witchcran. I do instruct it not to correct old-time spellings, and only to tidy up obvious mistakes. I guess it got confused on this one
Great idea. This should be fairly easy to do with the embedding vectors I have for the semantic search, using some clustering tools. Adding it to my backlog now!
Hmm, I just tried to upload the jpgs of some of todays samples to Arcanum via https://www.arcanum.com/en/newspaper-segmentation/try-it/ and it didn't work. I'll try it again later, but it seems based on a cursory look that it wouldn't return info that I would need to correct it if I didn't like the output, and that I'd still have to stitch the individual pages back together myself?
In hindsight it was probably a terrible example to you use, because people will think the OCR is off, but if you click on the clipping (or download the PDF from the PDF link at the top) and zoom in, you'll see that it's verbatim quoting some ancient text, which uses a lot of old-timey spelling (wickked e.g. is actually spelled wickked in the article), so I'm pretty happy with the quality it managed to eek out on that!
Check out the other examples for a more representative quality :-)
Yes I agree the layouts are the trickiest part. I tried a few and ended up using some of the Paddle Paddle models for document layout analysis and orientation and such, which give bounding boxes and predicted reading order, but the reading orders aren't great even with SOTA most recent models on complex layouts, or even simple layouts when you have mastheads or images or other artifacts to work around. It's still valuable information that can be combined with heuristics though to stitch together a more accurate reading order, as the starting point of a pipeline
Thank you, I really appreciate it. I will see if I can figure out how to do that, or like "if you're authed, you can try the Sleuth or get x free searches a month"? The balance is trying to do that without (potentially) overwhelming the databases, more so than trying intentionally to gate people out from anything. I'll figure it out!
I don't know if you looked at the "Label Specific" search, but I think I could fairly easily isolate that to a particular label and sub-type for people to search within without much risk to the backend. Any thoughts on a good category?
A few examples you can click on without having to authenticate or click the free trial (no cc if you do though and I won't bother you or chase you with spam etc...)