Comments on: Improving DeepSpell Code

By: Alex Skasgard

Alex Skasgard — Tue, 05 Feb 2019 16:03:00 +0000

Can you share new approach? @yaakov_belch:disqus

By: Yaakov Belch

Yaakov Belch — Wed, 13 Dec 2017 07:04:00 +0000

Nope. We are now using a different approach to spell checking.

By: Dheeraj

Dheeraj — Wed, 13 Dec 2017 06:02:00 +0000

Have you had any improvents in accuracy from 0.5?If so,What did you change in your code?

By: Yaakov Belch

Yaakov Belch — Sun, 09 Jul 2017 06:40:00 +0000

I also cannot replicate the success: Running the unchanged program on the recommended AWS EC2 image for 5 days: loss 0.16, acc: 0.5; none of the test samples is fully corrected. Some spelling mistakes are fixed — but other mistakes are introduced, often towards the end of the sentence.

By: surmenok

surmenok — Sun, 18 Jun 2017 01:56:00 +0000

Just did a benchmark, on GeForce 980 Ti, with 40 character inputs. Inference time depends on batch size.

Batch size 1: 30 ms per batch
Batch size 10: 35 ms per batch
Batch size 100: 309 ms per batch

By: Wolf Garbe

Wolf Garbe — Wed, 03 May 2017 09:33:00 +0000

Thank you very much for the reply. But what would be the correction time of “and who couQdn’tread in sixth grade” in milliseconds (the example from your comment)? I understand this also depends on the hardware, but I would really be interested in a raw figure/the order of magnitude to put the deep learning approach into perspective to classical approaches like Peter Norvig’s.

By: surmenok

surmenok — Wed, 03 May 2017 05:05:00 +0000

I haven’t analyzed quality of spelling correction on real documents yet. Performance of spelling correction should be linear function on input text length.

By: Wolf Garbe

Wolf Garbe — Fri, 28 Apr 2017 08:28:00 +0000

Could you share some insights on the correction performance? E.g. how many milliseconds does it take to correct “couQdn’tread” into “couldn’t read”? And how is the correction time dependent from the input text length (number of words) and edit distance (number of errors)? Constant/Linear/exponential?

By: surmenok

surmenok — Sun, 19 Mar 2017 23:08:00 +0000

I trained the model on a machine with GeForce 980Ti. Loss plateaued on the level about 0.2 after 3 days of training.
Test examples after 5 days of training:

Epoch 438:

Q Te threats of lawsuits againsit two…..
A The threats of lawsuits against two…..
Б≤▒ The threats of lawsuits against two…..
—
Q forward to working with Presidect…….
A forward to working with President…….
Б≤▒ forward to working with President…….
—
Q Black Warror River in Tuscaloosa……..
A Black Warrior River in Tuscaloosa…….
Б≤▓ Black Warror River in Tuscaloos.a…….
—
Q 13 and who couQdn’tread in sixth grade..
A 13 and who couldn’t read in sixth grade.
Б≤▒ 13 and who couldn’t read in sixth grade.

By: Matteo Tomassetti

Matteo Tomassetti — Fri, 24 Feb 2017 13:25:00 +0000

very nice blog post! I will definitely try you version because I’m having some problem with the original one.
Can I ask you if you were able to replicate Tal Weiss’s results? If so, which machine did you use and how long did it take you?