linguist/samples/SubRip Text/Adding.NCL.Language.S01E01.1080p.BluRay.x264.srt

1
00:00:01,250 --> 00:00:03,740
Adding NCL language.

2
00:00:04,600 --> 00:00:08,730
Thanks for the pull request! Do you know if these files are NCL too?

3
00:00:09,800 --> 00:00:13,700
Those are poorly-named documentation files for NCL functions.

4
00:00:14,560 --> 00:00:17,200
- What's better?
- This is better.

5
00:00:18,500 --> 00:00:23,000
- Would it be correct to recognise these files as text?
- Yes.

6
00:00:23,890 --> 00:00:30,000
In that case, could you add "NCL" to the text entry in languages.yml too?

7
00:00:30,540 --> 00:00:35,250
I added the example to "Text" and updated the license in the grammar submodule.

8
00:00:38,500 --> 00:00:42,360
Cloning the submodule fails for me in local with this URL.

9
00:00:42,360 --> 00:00:45,250
Could you use Git or HTTPS...?

10
00:00:46,810 --> 00:00:50,000
I updated the grammar submodule link to HTTPS.

11
00:00:51,100 --> 00:00:57,000
It's still failing locally. I don't think you can just update the .gitmodules file.

12
00:00:57,750 --> 00:01:03,000
You'll probably have to remove the submodule and add it again to be sure.

13
00:01:04,336 --> 00:01:11,800
- I'll see first if it's not an issue on my side...
- I removed the submodule and added it back with HTTPS.

14
00:01:13,670 --> 00:01:18,000
I tested the detection of NCL files with 2000 samples.

15
00:01:18,000 --> 00:01:25,000
The Bayesian classifier doesn't seem to be very good at distinguishing text from NCL.

16
00:01:25,000 --> 00:01:30,740
We could try to improve it by adding more samples, or we can define a new heuristic rule.

17
00:01:31,300 --> 00:01:36,200
- Do you want me to send you the sample files?
- Yes, please do.

18
00:01:37,500 --> 00:01:39,500
In your inbox.

19
00:01:41,285 --> 00:01:48,216
- So if I manually go through these and sort out the errors, would that help?
- Not really.

20
00:01:48,540 --> 00:01:55,145
It's a matter of keywords so there's not much to do there except for adding new samples.

21
00:01:55,447 --> 00:02:02,000
If adding a few more samples doesn't improve things, we'll see how to define a new heuristic rule.

22
00:02:04,740 --> 00:02:09,600
- I added quite a few NCL samples.
- That's a bit over the top, isn't it?

23
00:02:10,250 --> 00:02:16,000
We currently can't add too many samples because of #2117.

24
00:02:18,000 --> 00:02:20,830
(sigh) I decreased the number of added samples.

25
00:02:21,630 --> 00:02:25,300
Could you test the detection results in local with the samples I gave you?

26
00:02:26,000 --> 00:02:28,670
- What is the command to run that test?
- Here...

27
00:02:28,716 --> 00:02:38,650
[Coding intensifies]

28
00:02:38,650 --> 00:02:43,330
It is getting hung up on a false detection of Frege in one of the Text samples.

29
00:02:43,540 --> 00:02:46,115
Do you have any suggestions for implementing a heuristic?

30
00:02:47,640 --> 00:02:55,200
#2441 should fix this. In the meantime, you can change this in "test_heuristics.rb"

31
00:02:55,165 --> 00:02:57,240
Why did you have to change this?

32
00:02:57,777 --> 00:03:04,480
- It doesn't work for me unless I do that.
- Hum, same for me. Arfon, does it work for you?

33
00:03:04,920 --> 00:03:08,830
Requiring linguist/language doesn't work for me either.

34
00:03:09,300 --> 00:03:13,885
We restructured some of the requires a while ago and I think this is just out-of-date code.

35
00:03:14,065 --> 00:03:20,950
From a large sample of known NCL files taken from Github, it's now predicting with about 98% accuracy.

36
00:03:21,183 --> 00:03:28,000
For a large sample of other files with the NCL extension, it is around 92%.

37
00:03:27,880 --> 00:03:30,950
From those, nearly all of the errors come from one GitHub repository,

38
00:03:30,950 --> 00:03:34,160
and they all contain the text strings, "The URL" and "The Title".

39
00:03:35,660 --> 00:03:43,260
- Do you mean 92% files correctly identified as text?
- Yes, it correctly identifies 92% as text.

40
00:03:44,000 --> 00:03:46,150
I'd really like to see this dramatically reduced.

41
00:03:46,150 --> 00:03:51,150
What happens if we reduce to around 5 NCL sample files?

42
00:03:51,150 --> 00:03:52,600
Does Linguist still do a reasonable job?

43
00:03:53,470 --> 00:03:58,190
I reduced it to 16 NCL samples and 8 text samples.

44
00:03:58,190 --> 00:04:01,720
It correctly classifies my whole set of known NCL files.

45
00:04:01,870 --> 00:04:05,730
I tried with 5 samples but could not get the same level of accuracy.

46
00:04:06,670 --> 00:04:10,400
It incorrectly classifies all of the NCL files in this GitHub repository.

47
00:04:11,130 --> 00:04:14,660
All of these files contain the text strings, "THE_URL:" and "THE_TITLE:".

48
00:04:14,660 --> 00:04:19,500
It did not misclassify any other text-files with the extension NCL.

49
00:04:19,970 --> 00:04:25,188
With 100% accuracy? Does that mean it that the results are better with less samples??

50
00:04:25,610 --> 00:04:31,190
I also removed a sample text-file which should have been classified as an NCL file.

51
00:04:31,000 --> 00:04:35,895
I think that probably made most of the difference, although I didn't test it atomically.

52
00:04:35,895 --> 00:04:38,370
Okay, that makes more sense.

53
00:04:39,515 --> 00:04:43,450
I don't get the same results for the text files. Full results here.

54
00:04:44,650 --> 00:04:50,000
They all look correctly classified to me, except for the ones in Fanghuan's repository.

55
00:04:50,000 --> 00:04:55,920
I manually went through all of the ones where I didn't already know based on the filename or the repository owner.

56
00:04:56,526 --> 00:05:00,000
[Presses button] It now correctly classifies all of my test files.

57
00:05:00,000 --> 00:05:05,970
R. Pavlick, thanks for this. These changes will be live in the next release of Linguist. In the next couple of weeks.

58
00:05:05,970 --> 00:05:07,450
Great! Thanks.