[NTLK] NewTNG 3.0
lsynt3 at beanblossom.in.us
Thu Dec 8 11:37:00 EST 2011
At 5:01 AM -0500 12/2/11, Reilly001os at aol.com wrote:
>Apple built the print recognizer and licensed the cursive recognizer from someone else.
Correct. And just FYI...
The first generation recognition that was so bad and the much, much better second generation "cursive" recognizer were from the same lineage--a Russian company called Paragraph. The "print" recognizer, that only appeared in the second generation, was developed in the Advanced Technology Group at Apple Computer.
Prior to the modern generation of tablets (started quite a few years back at this point) Microsoft had a recognizer whose origins I don't know, but it was pretty dreadful. But since the official tablet computers, Microsoft licensed the same Paragraph technology that was in later Newtons, and, I was told by someone who worked on it, combined it at the whole-word level with some in-house recognition technology. I tried it not long after it was introduced and was surprised at how good it was until I learned its pedigree. Back then it still had some language model problems, and was worse than the Newton's print recognizer at dates and times, even though it was quite good at normal words, but I presume they fixed that at some point. And it does handle cursive writing and connected characters quite well.
Mac OS X's Inkwell is a direct descendent of the Newton's print recognizer, with a slightly improved language model and re-trained, slightly better neural network character recognition. Its performance with high-end tablets is quite good, however a lot of people use it with low-end, low spatial- and temporal-resolution tablets, and with those the performance is not as good as on the Newton, despite being a slightly improved version of the same technology. GIGO. (Garbage in, garbage out.) It's conceivable Inkwell's neural nets could be retrained with a bunch of data from these low-end tablets to improve its accuracy with them, but I don't believe this ever happened, and my connection to the project ended in early 2006.
More information about the NewtonTalk