We had to fix the following issues before running the algorithm and will apply relevant penalties during the final scoring: - invalid news output format, fixed extra comma; - invalid threads and top output format, fixed unescaped qoute
Bossy GnuDec 13, 2019 at 13:51
> invalid news output format, fixed extra comma; Fixed. Can be reproduced with only one language in a dataset (missed in my tests). > invalid threads and top output format, fixed unescaped qoute Fixed. OMG, shame on me! :)
Top threads of "Main" (both ru and en) consist of very loosely related articles.
Bossy GnuDec 13, 2019 at 12:45
Thank you for the comment. Yes, due to the extremely limited time I did not manage to configure the clustering algorithm perfectly. I used unmodified Chinese Whispers algotithm. And there is the well known problem - an object similar to an object which is similar to another object. This issue is fixed now. Any way IMO there are no any significant errors/problems in my implementation of the contest tasks.
Quite impressive news categorization, and news/no-news filtering! Not so great thread grouping though, but to me approach to improve that is rather clear. Processing speed is really great!
Bossy GnuDec 17, 2019 at 23:16
Thanks! "threads" & "top" are tuned and look nice now. I am going to upload new output jsons to show how it works now. I used quantized embeding models to keep my submission below 200MB. On the other hand it takes about 8-10 seconds for each language to restore the model while an uncopressed model can be loaded in a few tens of microseconds. And sure, you can load language models only once and use them in a further tasks. But I have to load and restore my model for each contest step (excluding the first step - languages detection).
Nobody added any issues yet...
Log in here to leave comments and take part in the Telegram Competitions. Please enter your phone number in the international format and we will send a confirmation message to your account via Telegram.