1
00:00:02,400 --> 00:00:07,160
Data visualizations are a powerful tool 
for communicating and exploring data.

2
00:00:07,160 --> 00:00:10,440
However, for blind or low vision 
users most data visualizations  

3
00:00:10,440 --> 00:00:14,720
are unusable because they are 
not designed to support them.

4
00:00:14,720 --> 00:00:18,760
A simple remedy is to provide text 
descriptions of visualizations.

5
00:00:18,760 --> 00:00:23,720
This however is frequently omitted 
especially for complex scientific charts.

6
00:00:23,720 --> 00:00:26,800
In this paper we introduce a 
method to automatically generate  

7
00:00:26,800 --> 00:00:31,160
text descriptions for upset. UpSet is 
a visualization technique for set data  

8
00:00:31,160 --> 00:00:34,680
that is commonly used in scientific publications.

9
00:00:34,680 --> 00:00:39,880
In our paper we analyze the types of patterns 
that are frequently found in upset plots and  

10
00:00:39,880 --> 00:00:44,520
develop a library that detects these patterns 
and generates text descriptions for them.

11
00:00:44,520 --> 00:00:46,880
In this video, however, we want to show you how  

12
00:00:46,880 --> 00:00:50,440
text descriptions are integrated 
into our web-based upset tool.

13
00:00:50,440 --> 00:00:54,360
It leverages our library to 
dynamically generate text descriptions.

14
00:00:54,360 --> 00:00:59,280
We will be using a dataset about movies for 
the purpose of this demonstration. This data  

15
00:00:59,280 --> 00:01:08,680
set shows the overlap between movie genres. We can 
change sorting or add and remove sets to the plot.

16
00:01:08,680 --> 00:01:12,360
We can reveal the text description 
in a side panel. We initially see  

17
00:01:12,360 --> 00:01:16,200
a short description which shows 
important at-a-glance information,  

18
00:01:16,200 --> 00:01:21,400
such as a description of the data set 
and the most important intersections.

19
00:01:21,400 --> 00:01:23,840
For more in-depth information, we can display a  

20
00:01:23,840 --> 00:01:28,320
long description. This longer form of 
the text description contains multiple  

21
00:01:28,320 --> 00:01:33,420
subsections which are all navigable 
with a keyboard and a screen reader.

22
00:01:33,420 --> 00:01:40,400
VoiceOver: Introduction. Heading level 
3. Upset introduction. This is an upset  

23
00:01:40,400 --> 00:01:43,360
plot that shows the intersections of six sets.

24
00:01:43,360 --> 00:01:50,360
Heading level 3. Data set properties. The 
data set contains 17 sets and 6,303 elements  

25
00:01:50,360 --> 00:01:55,480
of which six sets are shown in the plot.
Heading level 3. Set properties. List 2  

26
00:01:55,480 --> 00:02:02,400
items. Bullet. The largest set is action with 
503 elements followed by thriller with 492.

27
00:02:03,000 --> 00:02:07,360
Heading level 3. Intersection properties. 
List three items. Bullet. The plot is sorted  

28
00:02:07,360 --> 00:02:10,600
by size and descending order.
Bullet. There are 28 non-empty  

29
00:02:10,600 --> 00:02:15,880
intersections all of which are shown 
in the plot. Clickable. Voice over off.

30
00:02:15,880 --> 00:02:19,920
Our text descriptions are meant to be 
a template to be improved by humans. 

31
00:02:19,920 --> 00:02:24,560
We can provide a title and a caption to tell 
a story about the data set, but also provide  

32
00:02:24,560 --> 00:02:29,240
information about the data set that we can 
leverage to improve our text descriptions.

33
00:02:29,240 --> 00:02:35,080
The text description itself provides details 
about the data set, the properties of the sets,  

34
00:02:35,080 --> 00:02:39,520
and the properties of the intersections. 
We also provide higher level information  

35
00:02:39,520 --> 00:02:46,280
in the statistical information 
and trend analysis sections.

36
00:02:46,280 --> 00:02:51,040
Note that the text description updates 
as we change the upset plot. For example,  

37
00:02:51,040 --> 00:02:59,840
when we remove a set or change the 
sort order, the text is updated.

38
00:03:02,920 --> 00:03:08,480
Additionally, the generated text can be 
edited and saved. This enables users to  

39
00:03:08,480 --> 00:03:13,320
leverage the generated text as a baseline 
for them to provide additional insights,  

40
00:03:13,320 --> 00:03:23,120
modify generated stats, or simply 
create their own text description.

41
00:03:23,120 --> 00:03:27,480
Oour text generation is available as a 
language-independent web API, as well  

42
00:03:27,480 --> 00:03:33,480
as a python package. For example,we can generate 
text descriptions for a Python version of upset. 

43
00:03:33,480 --> 00:03:46,560
Here we generate an upset plot 
using the UpSetPlot python package.

44
00:03:46,560 --> 00:03:49,920
We can retrieve text descriptions 
in markdown format by simply calling  

45
00:03:49,920 --> 00:03:54,680
one function after creating the upset object.

46
00:03:54,680 --> 00:03:56,800
This text can now be customized and,  

47
00:03:56,800 --> 00:04:01,200
for example, included as the text 
description for a figure in a paper.

48
00:04:01,200 --> 00:04:05,280
To learn more about the validation of 
our approach please check out the paper.

49
00:04:05,280 --> 00:04:06,840
Thank you for watching our video.