large characters with ns greater than 9

[sounakdey]

Hi, congratulation on an amazing work... I understand having large characters with a number of parses more than 9 can make the factorial huge... do you have any suggestions on how to incorporate large characters?

[rfeinman]

Hi Souank - thank you for the kind remarks and for your interest in our work.

The use of random search during inference (step 2 in Appendix B) was a fill-in, and there are a number of better alternatives for large characters. With an auto-regressive model like ours—where P(stroke1, stroke2, stroke3) = P(stroke1)P(stroke2 | stroke1)P(stroke3 | stroke1, stroke2)—one option is to use "greedy" search, where you select parses based on P(stroke1), then refine the selection based on P(stroke2 | stroke1), etc. Beyond greedy search, a better alternative would be to use MCMC or particle filtering so that you are keeping some of the lower-scoring candidates with nonzero probability for future evaluations. Hope this is helpful, but I'd be happy to clarify further.