WEBVTT

1
00:00:00.000 --> 00:00:04.380
Hi, I'm Gustaf Gren from Stockholm University.

2
00:00:04.380 --> 00:00:07.520
So for quantitative sign language typological studies,

3
00:00:07.520 --> 00:00:09.660
we need data.

4
00:00:09.660 --> 00:00:13.400
But annotated data, especially those regarding non-manuals,

5
00:00:13.400 --> 00:00:16.040
like head shakes, is scarce.

6
00:00:16.040 --> 00:00:19.120
So what we analyze is whether or not

7
00:00:19.120 --> 00:00:22.240
we can use a neural-based pipeline

8
00:00:22.240 --> 00:00:25.000
to automatically identify head shakes

9
00:00:25.000 --> 00:00:30.600
for the purpose of basically trying to figure out

10
00:00:30.600 --> 00:00:36.680
the impact it has on annotation work.

11
00:00:36.680 --> 00:00:38.920
And while we find kind of limited performance

12
00:00:38.920 --> 00:00:42.560
on using this for automatic annotation,

13
00:00:42.560 --> 00:00:45.760
we find that these systems basically

14
00:00:45.760 --> 00:00:48.720
allow us to reduce the annotation burden by a lot.

15
00:00:48.720 --> 00:00:50.360
So the amount of frames that an annotator

16
00:00:50.360 --> 00:00:52.600
would have to go through.

17
00:00:52.600 --> 00:00:55.520
So this creates kind of a new potential cheaper

18
00:00:55.520 --> 00:01:00.000
way of annotating head shakes for sign language

19
00:01:00.000 --> 00:01:04.360
corpora, where these models would help the annotator,

20
00:01:04.360 --> 00:01:08.960
but not automatically identify head shakes.

21
00:01:08.960 --> 00:01:11.480
That's a very brief overview, of course.

22
00:01:11.480 --> 00:01:15.040
So check out our poster, come by.

23
00:01:15.040 --> 00:01:15.720
Yeah, thanks.