Kalika Bali: The giant leaps in language technology -- and who's left behind | TED

54,094 views ・ 2021-04-26

TED


Please double-click on the English subtitles below to play the video.

00:00
Transcriber:
0
0
7000
00:12
I'm Kalika Bali, I'm a linguist by training
1
12430
2800
00:15
and a technologist by profession,
2
15263
2334
00:17
I have worked in academia,
3
17630
1667
00:19
in startups, in small companies and multinationals for over two decades,
4
19330
5267
00:24
doing research in and building language technology systems.
5
24630
3733
00:28
My dream is to see technology work across the language barrier.
6
28363
4667
00:33
As a researcher at Microsoft Research Labs India
7
33030
3100
00:36
I work in the field of language technology and speech technology.
8
36163
5534
00:41
And I worry about how can we make technology accessible
9
41697
3600
00:45
to people across the board,
10
45330
2200
00:47
you know, irrespective of the language that they speak.
11
47530
2967
00:51
So natural language processing,
12
51663
1600
00:53
artificial intelligence, speech technology,
13
53297
2066
00:55
these are very big words, they are buzzwords right now.
14
55363
2600
00:57
Everybody is talking about what exactly is NLP or natural language processing.
15
57997
5533
01:03
So in very simple terms,
16
63530
1533
01:05
this is the part of computer science engineering
17
65063
3367
01:08
that makes machines process,
18
68463
2834
01:11
understand and generate natural language,
19
71330
2967
01:14
which is the language that humans speak.
20
74330
2233
01:17
When you are interacting with a bot trying to book your train tickets
21
77697
4700
01:22
or flight tickets,
22
82430
1267
01:23
when you are speaking to a voice-based digital assistant in your phone,
23
83697
4833
01:28
it's natural language processing
24
88530
1600
01:30
that underpins the entire technology that makes that work.
25
90163
3534
01:34
But how does this work?
26
94797
1266
01:36
How does NLP work?
27
96063
1567
01:37
In a very, very basic way,
28
97663
3900
01:41
it's about data.
29
101563
1800
01:43
So a huge amount of data of how actually humans use language
30
103363
6334
01:49
is then processed by certain algorithms and techniques
31
109697
5166
01:54
that make the machines learn the patterns
32
114863
2767
01:57
of natural language of humans, right?
33
117663
3700
02:01
These days, another buzzword that you hear a lot about is deep neural networks.
34
121363
5367
02:06
And these are the advanced techniques
35
126763
2467
02:09
that underpin a lot of the NLP stuff that happens right now.
36
129263
4134
02:13
And I will not go into the details of how that works,
37
133430
3333
02:16
but the thing that you really have to understand and keep in mind
38
136797
3400
02:20
is that all of this requires a humungous amount of data,
39
140197
5000
02:25
natural language data.
40
145197
1666
02:26
If you want a speech system to converse with you in Gujarati,
41
146863
5267
02:32
the first thing you require
42
152163
1367
02:33
is a lot of data of Gujarati people speaking to each other
43
153530
4533
02:38
in their own language.
44
158063
1867
02:41
So 2017, Microsoft came up with a speech recognition system
45
161663
4734
02:46
which was able to transcribe speech into text
46
166430
3600
02:50
better than a human did.
47
170030
1733
02:52
And this system was trained
48
172563
3167
02:55
on 200 million transcribed words.
49
175763
2667
02:58
In 2018, an English-Chinese machine translation system
50
178463
3967
03:02
was able to translate from English to Chinese
51
182463
2800
03:05
as well as any human bilingual could.
52
185297
2766
03:08
And this was trained on 18 million bilingual sentence pairs.
53
188063
4934
03:14
This is a very, very exciting time in natural language processing
54
194330
4433
03:18
and in technology as such.
55
198797
1466
03:20
You know, we are seeing science fiction, which we had read about and watched,
56
200297
4466
03:24
kind of come true in front of our own eyes.
57
204797
2833
03:27
We are making giant leaps in technical advancement.
58
207663
4467
03:32
But these giant leaps are limited to very few languages.
59
212163
6234
03:38
So Monojit Choudhury,
60
218430
1300
03:39
who's like a very good friend of mine
61
219763
1867
03:41
and a colleague,
62
221663
2134
03:43
he has studied this in some detail
63
223830
1733
03:45
and he has looked at resource distribution across languages in the world.
64
225563
4067
03:49
And he says that these follow what is called a power-law distribution,
65
229663
4100
03:53
which essentially means that there are four languages,
66
233797
2766
03:56
Arabic, Chinese, English and Spanish,
67
236563
2700
03:59
which have the maximum amount of resources available.
68
239297
3766
04:03
There are another handful of languages which can also benefit from, you know,
69
243063
5167
04:08
the resources and the technology that's available right now.
70
248263
3834
04:12
But there are 90 percent of the world's languages
71
252130
4833
04:16
which have no resources
72
256997
1800
04:18
or very little resources available.
73
258830
2067
04:20
This revolution that we are talking about
74
260930
2667
04:23
has essentially bypassed 5,000 languages of the world.
75
263630
4100
04:27
Now, what this means is that resource-rich languages
76
267763
2534
04:30
have technologies built for them,
77
270330
1800
04:32
so researchers and technologists get attracted towards them.
78
272163
3267
04:35
They build more technologies for them. They create more resources.
79
275463
3500
04:38
So it's like a rich getting richer kind of a cycle.
80
278997
2800
04:41
And the resource-poor languages stay poor,
81
281830
2400
04:44
there's no technology for them, nobody works for them.
82
284263
2600
04:46
And this divide, digital divide between languages
83
286863
3400
04:50
is ever-expanding
84
290297
1500
04:51
and by implication also the divide between the communities
85
291830
4633
04:56
that speak these languages is expanding.
86
296497
2500
05:00
So in Microsoft, in Project Ellora, we aim to bridge this gap.
87
300763
4767
05:06
We are trying to see how can we create more data by innovative methods,
88
306663
5567
05:12
have more techniques to build technology without having a lot of resources,
89
312263
5800
05:18
and what are the applications that can truly benefit these communities.
90
318063
4200
05:23
So at the moment, this might seem very theoretical,
91
323463
3334
05:26
like what is he talking about, data and techniques and technology.
92
326830
3133
05:29
So let me give you a very concrete example here.
93
329997
3066
05:33
I'm a linguist at heart, I love languages, and that's what I love talking about.
94
333063
5300
05:38
So let me tell you about a language that many of you might not know about.
95
338363
4367
05:42
Gondi.
96
342763
1267
05:44
Gondi is a South-Central Dravidian language.
97
344030
2700
05:46
It is spoken by three million people in five states of India.
98
346763
4434
05:51
And to put this in some kind of perspective,
99
351197
3000
05:54
Norwegian is spoken by five million people
100
354197
2833
05:57
and Welsh by a little under a million.
101
357030
2933
05:59
So Gondi is actually a pretty robust and pretty large community
102
359997
6200
06:06
of the Gond tribals in India.
103
366197
2900
06:09
But by UNESCO's Atlas of Languages in Danger,
104
369130
5067
06:14
Gondi is designated vulnerable status.
105
374197
4366
06:19
CGNet Swara is an NGO that provides a citizen journalism portal
106
379030
4533
06:23
for the Gond community
107
383563
1867
06:25
by making local stories accessible through mobile phones.
108
385463
4234
06:29
There's absolutely no tech support for Gondi.
109
389697
2933
06:32
There is no data available for Gondi, no resources available for Gondi.
110
392663
4800
06:37
So all content that is created, moderated and edited is done manually.
111
397497
5066
06:42
Now, under Project Ellora,
112
402563
2067
06:44
what we did was that we brought together all the stakeholders,
113
404663
2967
06:47
an NGOs like CGNet Swara,
114
407663
1800
06:49
and academic institutions, like IIIT Naya Raipur,
115
409497
3366
06:52
a not-for-profit children's book publisher,
116
412863
2200
06:55
like Pratham Books,
117
415063
1300
06:56
and most importantly, the speakers of the community.
118
416363
2434
06:58
The Gond tribals themselves participated in this activity
119
418830
4933
07:03
and for the first time edited and translated children’s books in Gondi.
120
423797
5633
07:09
We were able to put out 200 books for the very first time in Gondi,
121
429463
5334
07:14
so that the children had access to stories and books in their own language.
122
434830
4700
07:19
Another extension of this was Adivasi Radio,
123
439530
2267
07:21
which was like an app that we built and developed in Microsoft Research,
124
441830
4033
07:25
and then put out there, along with our stakeholders,
125
445863
4567
07:30
which takes a Hindi text-to-speech system
126
450463
3067
07:33
and allows it to read out news and articles provided by CGNet Swara
127
453530
6400
07:39
in Gondi language.
128
459963
2434
07:42
Users can now use this app to read,
129
462430
2767
07:45
watch news and access any information
130
465197
3433
07:48
through text and voice in their own language.
131
468663
3900
07:52
A very interesting thing is that this app is now being used to translate --
132
472563
3634
07:56
by the community to translate text from Hindi to Gondi.
133
476197
4833
08:01
Now, what that will result in is a lot of parallel data,
134
481030
3333
08:04
that we call parallel data,
135
484363
1534
08:05
that will allow us to build machine translation systems for Gondi,
136
485930
3500
08:09
which will truly open up a window for the Gond community to the world.
137
489463
5900
08:15
And what is even more important is now we know how to do this.
138
495363
3567
08:18
We have the entire pipeline and we can replicate this for any language
139
498963
4600
08:23
and any language community
140
503563
1900
08:25
which is in a similar situation as the Gond tribals.
141
505497
3033
08:29
Also education -- yes, you know, information access -- yes,
142
509630
4833
08:34
but what about earning a living?
143
514497
2700
08:37
Right? What about -- how can we make these people earn a living
144
517830
4200
08:42
through the digital tools that all of us just take for granted these days?
145
522030
3867
08:45
Vivek Seshadri, who's another researcher at MSR,
146
525930
2533
08:48
and his collaborator, Manu Chopra,
147
528497
2100
08:50
they've designed a platform called Karya
148
530630
2500
08:53
for providing digital microtasks to the underserved communities.
149
533163
4600
08:57
His aim was basically to find a way to provide a means of dignified labor
150
537797
5433
09:03
to the populations, the rural populations
151
543263
2034
09:05
and the urban poor populations of this country.
152
545330
2300
09:08
They don't have access to all the knowledge
153
548530
3000
09:11
to use the digital platforms
154
551530
2567
09:14
that all of us use every day without even thinking, right?
155
554130
4200
09:18
But ...
156
558930
1233
09:20
Here is a large
157
560863
2434
09:23
literate population that wants to work, right,
158
563330
4133
09:27
and how can we make this possible for them?
159
567497
2766
09:30
So Karya is one such way
160
570297
3500
09:33
through which this population can get on to the digital world
161
573830
4033
09:37
and, you know,
162
577863
1434
09:39
through that find work and do tasks that can then earn them money.
163
579330
4433
09:43
So we saw this and we thought, oh, this is wonderful.
164
583797
2500
09:46
We could probably use this for data collection as well.
165
586330
2600
09:48
So we went to Amale,
166
588963
2000
09:50
which is a small village of 200 people
167
590997
3266
09:54
in the Wada district of Maharashtra
168
594297
1933
09:56
and decided to use Karya to collect Marathi data.
169
596263
2667
09:58
Now, I know what you are thinking --
170
598963
1900
10:00
I'm sure a lot of Marathi speakers also in the audience --
171
600930
2767
10:03
that Marathi is not a low-resource language.
172
603697
2400
10:06
Marathi is definitely a mainstream language of the country.
173
606130
3600
10:09
But as far as language technology is concerned,
174
609763
2500
10:12
Marathi is a low-resource language.
175
612297
2333
10:14
So we went to this village
176
614663
1600
10:16
and we had a very successful data-collection trip.
177
616297
3800
10:20
And, you know, this village is very remote.
178
620130
3567
10:23
They have no TV, they have no electricity,
179
623697
2700
10:26
they have no mobile signal.
180
626430
3833
10:30
You have to climb a hill and wave your phone around
181
630297
2633
10:32
if you want to, you know, use your mobile to call anyone.
182
632963
3200
10:37
So they gave us all this data.
183
637197
1766
10:38
But more than that, they gave us very valuable lessons in life.
184
638997
3266
10:43
One is this pride in one's own language.
185
643063
3000
10:46
The people of Amale were thrilled to be doing this
186
646063
2667
10:48
because they were advancing their own language by doing this.
187
648763
5834
10:54
The second was the value of community.
188
654630
1833
10:56
Very quickly, this became a village community effort.
189
656497
3900
11:00
People would gather together in tasks and do this together as a group.
190
660430
4867
11:05
And the third is the importance of storytelling.
191
665330
3767
11:09
People of Amale were so starved of content that in the morning, during the daytime,
192
669130
6067
11:15
they would do recordings of stories in Karya
193
675197
4066
11:19
and then in the evening they would gather the entire village
194
679297
2833
11:22
and retell and recount these stories to the village.
195
682163
3834
11:27
So as scientists, we get so caught up
196
687563
2167
11:29
in the science and technology part of what we are doing, you know --
197
689763
3700
11:33
which is the next best model to have,
198
693497
1966
11:35
how can we increase the accuracy of my system,
199
695497
2933
11:38
how can I build the next best system there is --
200
698463
4600
11:43
that we forget the reason why we are doing this: the people.
201
703063
3700
11:46
And any successful technology is the one that keeps the people and the users
202
706797
5466
11:52
up front and center.
203
712297
1566
11:54
And when they start doing that,
204
714830
1533
11:56
we also realize that technology is probably a very small part of this
205
716363
3667
12:00
and there are other things in the story.
206
720030
2367
12:02
Maybe there are social, cultural and policy interventions
207
722430
3333
12:05
that are required, as much as technology.
208
725797
2200
12:09
So some time back, I worked on a project called VideoKheti
209
729030
3267
12:12
that allowed Hindi-speaking farmers in Central India
210
732330
3600
12:15
to search for agricultural videos by speaking into a phone-based app.
211
735963
6667
12:23
So we went to Madhya Pradesh to collect data for this,
212
743363
3434
12:26
and we came back and we were training our models
213
746830
2300
12:29
and we discovered we're getting very bad results.
214
749163
2367
12:31
This is not working.
215
751530
1267
12:32
So we were very confused. Why is this happening?
216
752830
2267
12:35
So we looked deeper and deeper into the data
217
755130
2267
12:37
and discovered that, yes, we had collected data
218
757430
2233
12:39
from what we thought was a very silent, quiet village in the evening.
219
759697
4666
12:44
But what we hadn't heard while we were doing this
220
764363
3400
12:47
was that there was this constant buzz of night insects, you know?
221
767797
4100
12:51
So throughout the recordings, we had this "bzz" of the insects,
222
771930
3467
12:55
which was actually distorting our speech.
223
775430
2533
12:58
The second thing was that when we went there
224
778797
2233
13:01
to kind of test our app in the village,
225
781030
3867
13:04
I and my colleague Indrani Medhi,
226
784930
2533
13:07
who is a very well-regarded design researcher,
227
787497
3600
13:11
we found that the women couldn't pronounce the sanskritized words
228
791130
4400
13:15
that we had for some of the search terms.
229
795530
2767
13:18
So, like ...
230
798330
1500
13:21
(speaks Hindi)
231
801663
2800
13:24
Which is like the term for chemical pesticides, right?
232
804497
3866
13:28
Because we got these terms from the agricultural extension center
233
808363
5534
13:33
and the women, even though they are farming,
234
813930
2100
13:36
do not interact with that center at all.
235
816030
2867
13:38
The men do, the women probably use something much simpler, like ...
236
818930
3967
13:42
(speaks Hindi)
237
822930
1300
13:44
Which basically means killing pests with medicine.
238
824263
3534
13:48
So what I have learned through my journey
239
828430
3867
13:52
and what I would like to put across to you --
240
832330
2900
13:55
by now, I hope you've understood me,
241
835263
2000
13:57
is that there is the majority of the world's languages
242
837297
3433
14:00
that require intensive investment for resource creation
243
840763
4267
14:05
if they are to benefit from language technology.
244
845030
2567
14:07
And this is unlikely to happen in a very fast and efficient manner.
245
847630
5367
14:13
So it is extremely important for us to ensure
246
853963
2934
14:16
that the community derives maximum benefit
247
856930
3533
14:20
from whatever that we are doing in the language tech area.
248
860497
3966
14:24
And to do this and deliver a positive social impact
249
864497
3466
14:27
on these communities,
250
867997
1466
14:29
we follow what we call the modified 4-D design thinking methodology.
251
869497
4733
14:34
So the 4-D means: discover, design, develop and deploy.
252
874263
5200
14:39
So discover the problem that language technology can solve
253
879497
3066
14:42
for a particular language community.
254
882563
2200
14:44
This observation-led approach can help allocate resources
255
884797
3233
14:48
where they are most needed,
256
888030
1700
14:49
designed for the users and their language,
257
889763
2767
14:52
understand the diversity in the linguistic properties
258
892530
3367
14:55
and the languages of the world.
259
895930
2100
14:58
And don't think, oh, this is made for English.
260
898030
2400
15:00
Now, how can we just adapt it for Marathi or for Gondi, right?
261
900463
4334
15:04
Develop rapidly and deploy frequently.
262
904830
2600
15:07
It's an iterative process that will help you fail fast
263
907463
3500
15:10
and early failures will eventually lead to success.
264
910997
3366
15:15
The important thing is to persevere.
265
915497
1966
15:17
Do not give up.
266
917497
1366
15:18
And I remember the story of these two Aborigine Australian women,
267
918863
5734
15:24
Patricia O'Connor and Ysola Best.
268
924630
3800
15:29
In the mid-90s, they went to the University of Queensland
269
929763
3134
15:32
and they wanted to learn their own language, called Yugambeh,
270
932930
3333
15:36
and they were told very bluntly, "Your language is dead.
271
936297
2633
15:38
It's been dead for three decades.
272
938963
1600
15:40
You cannot work on this. Find something else to work on."
273
940563
3867
15:44
They did not give up.
274
944463
1267
15:45
They went to the community,
275
945763
1600
15:47
they dug up oral memories, oral traditions, oral literature,
276
947363
4867
15:52
and founded the Yugambeh Museum,
277
952263
3367
15:55
which became the most important cultural and linguistic center for the language
278
955663
5434
16:01
and its community.
279
961130
1767
16:02
They did not have technology. They only had their willpower.
280
962930
4033
16:06
Now, with the power of technology,
281
966997
2233
16:09
we can ensure that the next page is written in Salmi from Finland,
282
969263
5767
16:15
Lillooet from Canada or Mundari from India.
283
975030
3467
16:19
Thank you.
284
979163
1000
About this website

This site will introduce you to YouTube videos that are useful for learning English. You will see English lessons taught by top-notch teachers from around the world. Double-click on the English subtitles displayed on each video page to play the video from there. The subtitles scroll in sync with the video playback. If you have any comments or requests, please contact us using this contact form.

https://forms.gle/WvT1wiN1qDtmnspy7