5 ms·
Funny The arms/anus confusion should be fixed with a language model on top of the letter predicting network.
by visarga 8y ago
Funny
The arms/anus confusion should be fixed with a language model on top of the letter predicting network.
- duskwuff 8y agoWithout a very long-range model I don't think that would help. "in his/her anus" and "in his/her arms" can both be correct in the right circumstances; it takes quite a bit of surrounding context to tell which one is more likely. (While doing some research in Google Books I even found a couple that looked like OCR errors until I read beyond the search snippet.)
- abecedarius 8y agoHow does OCR software integrate letter and language models? Do they first make a best guess at the letters and then try to correct it with the language model? https://en.wikipedia.org/wiki/Optical_character_recognition#Techniques https://en.wikipedia.org/wiki/Optical_character_recognition#... gives me that impression, but I'm not sure. Brains are said to have a lot of feedback from higher levels of sensory processing to lower. Maybe you don't need as good a language model if its evidence is integrated more tightly with the rest.
- amelius 8y agoBut that would completely ruin the article, if it were OCR'ed.
- mort96 8y agoOnly without human oversight. Human proofreading might not be economically feasible, but maybe detecting likely OCR errors and making humans decide whether it's an actual error or not would be feasible.