TEDxBoston 2011
Jean-Baptiste Michel + Erez Lieberman Aiden: What we learned from 5 million books
Filmed:
Readability: 3.9
2,049,453 views
Have you played with Google Labs' Ngram Viewer? It's an addicting tool that lets you search for words and ideas in a database of 5 million books from across centuries. Erez Lieberman Aiden and Jean-Baptiste Michel show us how it works, and a few of the surprising things we can learn from 500 billion words.
Jean-Baptiste Michel - Data researcher
Jean-Baptiste Michel looks at how we can use large volumes of data to better understand our world. Full bioErez Lieberman Aiden - Researcher
Erez Lieberman Aiden pursues a broad range of research interests, spanning genomics, linguistics, mathematics ... Full bio
Jean-Baptiste Michel looks at how we can use large volumes of data to better understand our world. Full bioErez Lieberman Aiden - Researcher
Erez Lieberman Aiden pursues a broad range of research interests, spanning genomics, linguistics, mathematics ... Full bio
Double-click the English transcript below to play the video.
00:15
Erez Lieberman Aiden: Everyone knows
0
0
2000
00:17
that a picture is worth a thousand words.
1
2000
3000
00:22
But we at Harvard
2
7000
2000
00:24
were wondering if this was really true.
3
9000
3000
00:27
(Laughter)
4
12000
2000
00:29
So we assembled a team of experts,
5
14000
4000
00:33
spanning Harvard, MIT,
6
18000
2000
00:35
The American Heritage Dictionary, The Encyclopedia Britannica
7
20000
3000
00:38
and even our proud sponsors,
8
23000
2000
00:40
the Google.
9
25000
3000
00:43
And we cogitated about this
10
28000
2000
00:45
for about four years.
11
30000
2000
00:47
And we came to a startling conclusion.
12
32000
5000
00:52
Ladies and gentlemen, a picture is not worth a thousand words.
13
37000
3000
00:55
In fact, we found some pictures
14
40000
2000
00:57
that are worth 500 billion words.
15
42000
5000
01:02
Jean-Baptiste Michel: So how did we get to this conclusion?
16
47000
2000
01:04
So Erez and I were thinking about ways
17
49000
2000
01:06
to get a big picture of human culture
18
51000
2000
01:08
and human history: change over time.
19
53000
3000
01:11
So many books actually have been written over the years.
20
56000
2000
01:13
So we were thinking, well the best way to learn from them
21
58000
2000
01:15
is to read all of these millions of books.
22
60000
2000
01:17
Now of course, if there's a scale for how awesome that is,
23
62000
3000
01:20
that has to rank extremely, extremely high.
24
65000
3000
01:23
Now the problem is there's an X-axis for that,
25
68000
2000
01:25
which is the practical axis.
26
70000
2000
01:27
This is very, very low.
27
72000
2000
01:29
(Applause)
28
74000
3000
01:32
Now people tend to use an alternative approach,
29
77000
3000
01:35
which is to take a few sources and read them very carefully.
30
80000
2000
01:37
This is extremely practical, but not so awesome.
31
82000
2000
01:39
What you really want to do
32
84000
3000
01:42
is to get to the awesome yet practical part of this space.
33
87000
3000
01:45
So it turns out there was a company across the river called Google
34
90000
3000
01:48
who had started a digitization project a few years back
35
93000
2000
01:50
that might just enable this approach.
36
95000
2000
01:52
They have digitized millions of books.
37
97000
2000
01:54
So what that means is, one could use computational methods
38
99000
3000
01:57
to read all of the books in a click of a button.
39
102000
2000
01:59
That's very practical and extremely awesome.
40
104000
3000
02:03
ELA: Let me tell you a little bit about where books come from.
41
108000
2000
02:05
Since time immemorial, there have been authors.
42
110000
3000
02:08
These authors have been striving to write books.
43
113000
3000
02:11
And this became considerably easier
44
116000
2000
02:13
with the development of the printing press some centuries ago.
45
118000
2000
02:15
Since then, the authors have won
46
120000
3000
02:18
on 129 million distinct occasions,
47
123000
2000
02:20
publishing books.
48
125000
2000
02:22
Now if those books are not lost to history,
49
127000
2000
02:24
then they are somewhere in a library,
50
129000
2000
02:26
and many of those books have been getting retrieved from the libraries
51
131000
3000
02:29
and digitized by Google,
52
134000
2000
02:31
which has scanned 15 million books to date.
53
136000
2000
02:33
Now when Google digitizes a book, they put it into a really nice format.
54
138000
3000
02:36
Now we've got the data, plus we have metadata.
55
141000
2000
02:38
We have information about things like where was it published,
56
143000
3000
02:41
who was the author, when was it published.
57
146000
2000
02:43
And what we do is go through all of those records
58
148000
3000
02:46
and exclude everything that's not the highest quality data.
59
151000
4000
02:50
What we're left with
60
155000
2000
02:52
is a collection of five million books,
61
157000
3000
02:55
500 billion words,
62
160000
3000
02:58
a string of characters a thousand times longer
63
163000
2000
03:00
than the human genome --
64
165000
3000
03:03
a text which, when written out,
65
168000
2000
03:05
would stretch from here to the Moon and back
66
170000
2000
03:07
10 times over --
67
172000
2000
03:09
a veritable shard of our cultural genome.
68
174000
4000
03:13
Of course what we did
69
178000
2000
03:15
when faced with such outrageous hyperbole ...
70
180000
3000
03:18
(Laughter)
71
183000
2000
03:20
was what any self-respecting researchers
72
185000
3000
03:23
would have done.
73
188000
3000
03:26
We took a page out of XKCD,
74
191000
2000
03:28
and we said, "Stand back.
75
193000
2000
03:30
We're going to try science."
76
195000
2000
03:32
(Laughter)
77
197000
2000
03:34
JM: Now of course, we were thinking,
78
199000
2000
03:36
well let's just first put the data out there
79
201000
2000
03:38
for people to do science to it.
80
203000
2000
03:40
Now we're thinking, what data can we release?
81
205000
2000
03:42
Well of course, you want to take the books
82
207000
2000
03:44
and release the full text of these five million books.
83
209000
2000
03:46
Now Google, and Jon Orwant in particular,
84
211000
2000
03:48
told us a little equation that we should learn.
85
213000
2000
03:50
So you have five million, that is, five million authors
86
215000
3000
03:53
and five million plaintiffs is a massive lawsuit.
87
218000
3000
03:56
So, although that would be really, really awesome,
88
221000
2000
03:58
again, that's extremely, extremely impractical.
89
223000
3000
04:01
(Laughter)
90
226000
2000
04:03
Now again, we kind of caved in,
91
228000
2000
04:05
and we did the very practical approach, which was a bit less awesome.
92
230000
3000
04:08
We said, well instead of releasing the full text,
93
233000
2000
04:10
we're going to release statistics about the books.
94
235000
2000
04:12
So take for instance "A gleam of happiness."
95
237000
2000
04:14
It's four words; we call that a four-gram.
96
239000
2000
04:16
We're going to tell you how many times a particular four-gram
97
241000
2000
04:18
appeared in books in 1801, 1802, 1803,
98
243000
2000
04:20
all the way up to 2008.
99
245000
2000
04:22
That gives us a time series
100
247000
2000
04:24
of how frequently this particular sentence was used over time.
101
249000
2000
04:26
We do that for all the words and phrases that appear in those books,
102
251000
3000
04:29
and that gives us a big table of two billion lines
103
254000
3000
04:32
that tell us about the way culture has been changing.
104
257000
2000
04:34
ELA: So those two billion lines,
105
259000
2000
04:36
we call them two billion n-grams.
106
261000
2000
04:38
What do they tell us?
107
263000
2000
04:40
Well the individual n-grams measure cultural trends.
108
265000
2000
04:42
Let me give you an example.
109
267000
2000
04:44
Let's suppose that I am thriving,
110
269000
2000
04:46
then tomorrow I want to tell you about how well I did.
111
271000
2000
04:48
And so I might say, "Yesterday, I throve."
112
273000
3000
04:51
Alternatively, I could say, "Yesterday, I thrived."
113
276000
3000
04:54
Well which one should I use?
114
279000
3000
04:57
How to know?
115
282000
2000
04:59
As of about six months ago,
116
284000
2000
05:01
the state of the art in this field
117
286000
2000
05:03
is that you would, for instance,
118
288000
2000
05:05
go up to the following psychologist with fabulous hair,
119
290000
2000
05:07
and you'd say,
120
292000
2000
05:09
"Steve, you're an expert on the irregular verbs.
121
294000
3000
05:12
What should I do?"
122
297000
2000
05:14
And he'd tell you, "Well most people say thrived,
123
299000
2000
05:16
but some people say throve."
124
301000
3000
05:19
And you also knew, more or less,
125
304000
2000
05:21
that if you were to go back in time 200 years
126
306000
3000
05:24
and ask the following statesman with equally fabulous hair,
127
309000
3000
05:27
(Laughter)
128
312000
3000
05:30
"Tom, what should I say?"
129
315000
2000
05:32
He'd say, "Well, in my day, most people throve,
130
317000
2000
05:34
but some thrived."
131
319000
3000
05:37
So now what I'm just going to show you is raw data.
132
322000
2000
05:39
Two rows from this table of two billion entries.
133
324000
4000
05:43
What you're seeing is year by year frequency
134
328000
2000
05:45
of "thrived" and "throve" over time.
135
330000
3000
05:49
Now this is just two
136
334000
2000
05:51
out of two billion rows.
137
336000
3000
05:54
So the entire data set
138
339000
2000
05:56
is a billion times more awesome than this slide.
139
341000
3000
05:59
(Laughter)
140
344000
2000
06:01
(Applause)
141
346000
4000
06:05
JM: Now there are many other pictures that are worth 500 billion words.
142
350000
2000
06:07
For instance, this one.
143
352000
2000
06:09
If you just take influenza,
144
354000
2000
06:11
you will see peaks at the time where you knew
145
356000
2000
06:13
big flu epidemics were killing people around the globe.
146
358000
3000
06:16
ELA: If you were not yet convinced,
147
361000
3000
06:19
sea levels are rising,
148
364000
2000
06:21
so is atmospheric CO2 and global temperature.
149
366000
3000
06:24
JM: You might also want to have a look at this particular n-gram,
150
369000
3000
06:27
and that's to tell Nietzsche that God is not dead,
151
372000
3000
06:30
although you might agree that he might need a better publicist.
152
375000
3000
06:33
(Laughter)
153
378000
2000
06:35
ELA: You can get at some pretty abstract concepts with this sort of thing.
154
380000
3000
06:38
For instance, let me tell you the history
155
383000
2000
06:40
of the year 1950.
156
385000
2000
06:42
Pretty much for the vast majority of history,
157
387000
2000
06:44
no one gave a damn about 1950.
158
389000
2000
06:46
In 1700, in 1800, in 1900,
159
391000
2000
06:48
no one cared.
160
393000
3000
06:52
Through the 30s and 40s,
161
397000
2000
06:54
no one cared.
162
399000
2000
06:56
Suddenly, in the mid-40s,
163
401000
2000
06:58
there started to be a buzz.
164
403000
2000
07:00
People realized that 1950 was going to happen,
165
405000
2000
07:02
and it could be big.
166
407000
2000
07:04
(Laughter)
167
409000
3000
07:07
But nothing got people interested in 1950
168
412000
3000
07:10
like the year 1950.
169
415000
3000
07:13
(Laughter)
170
418000
3000
07:16
People were walking around obsessed.
171
421000
2000
07:18
They couldn't stop talking
172
423000
2000
07:20
about all the things they did in 1950,
173
425000
3000
07:23
all the things they were planning to do in 1950,
174
428000
3000
07:26
all the dreams of what they wanted to accomplish in 1950.
175
431000
5000
07:31
In fact, 1950 was so fascinating
176
436000
2000
07:33
that for years thereafter,
177
438000
2000
07:35
people just kept talking about all the amazing things that happened,
178
440000
3000
07:38
in '51, '52, '53.
179
443000
2000
07:40
Finally in 1954,
180
445000
2000
07:42
someone woke up and realized
181
447000
2000
07:44
that 1950 had gotten somewhat passé.
182
449000
4000
07:48
(Laughter)
183
453000
2000
07:50
And just like that, the bubble burst.
184
455000
2000
07:52
(Laughter)
185
457000
2000
07:54
And the story of 1950
186
459000
2000
07:56
is the story of every year that we have on record,
187
461000
2000
07:58
with a little twist, because now we've got these nice charts.
188
463000
3000
08:01
And because we have these nice charts, we can measure things.
189
466000
3000
08:04
We can say, "Well how fast does the bubble burst?"
190
469000
2000
08:06
And it turns out that we can measure that very precisely.
191
471000
3000
08:09
Equations were derived, graphs were produced,
192
474000
3000
08:12
and the net result
193
477000
2000
08:14
is that we find that the bubble bursts faster and faster
194
479000
3000
08:17
with each passing year.
195
482000
2000
08:19
We are losing interest in the past more rapidly.
196
484000
5000
08:24
JM: Now a little piece of career advice.
197
489000
2000
08:26
So for those of you who seek to be famous,
198
491000
2000
08:28
we can learn from the 25 most famous political figures,
199
493000
2000
08:30
authors, actors and so on.
200
495000
2000
08:32
So if you want to become famous early on, you should be an actor,
201
497000
3000
08:35
because then fame starts rising by the end of your 20s --
202
500000
2000
08:37
you're still young, it's really great.
203
502000
2000
08:39
Now if you can wait a little bit, you should be an author,
204
504000
2000
08:41
because then you rise to very great heights,
205
506000
2000
08:43
like Mark Twain, for instance: extremely famous.
206
508000
2000
08:45
But if you want to reach the very top,
207
510000
2000
08:47
you should delay gratification
208
512000
2000
08:49
and, of course, become a politician.
209
514000
2000
08:51
So here you will become famous by the end of your 50s,
210
516000
2000
08:53
and become very, very famous afterward.
211
518000
2000
08:55
So scientists also tend to get famous when they're much older.
212
520000
3000
08:58
Like for instance, biologists and physics
213
523000
2000
09:00
tend to be almost as famous as actors.
214
525000
2000
09:02
One mistake you should not do is become a mathematician.
215
527000
3000
09:05
(Laughter)
216
530000
2000
09:07
If you do that,
217
532000
2000
09:09
you might think, "Oh great. I'm going to do my best work when I'm in my 20s."
218
534000
3000
09:12
But guess what, nobody will really care.
219
537000
2000
09:14
(Laughter)
220
539000
3000
09:17
ELA: There are more sobering notes
221
542000
2000
09:19
among the n-grams.
222
544000
2000
09:21
For instance, here's the trajectory of Marc Chagall,
223
546000
2000
09:23
an artist born in 1887.
224
548000
2000
09:25
And this looks like the normal trajectory of a famous person.
225
550000
3000
09:28
He gets more and more and more famous,
226
553000
4000
09:32
except if you look in German.
227
557000
2000
09:34
If you look in German, you see something completely bizarre,
228
559000
2000
09:36
something you pretty much never see,
229
561000
2000
09:38
which is he becomes extremely famous
230
563000
2000
09:40
and then all of a sudden plummets,
231
565000
2000
09:42
going through a nadir between 1933 and 1945,
232
567000
3000
09:45
before rebounding afterward.
233
570000
3000
09:48
And of course, what we're seeing
234
573000
2000
09:50
is the fact Marc Chagall was a Jewish artist
235
575000
3000
09:53
in Nazi Germany.
236
578000
2000
09:55
Now these signals
237
580000
2000
09:57
are actually so strong
238
582000
2000
09:59
that we don't need to know that someone was censored.
239
584000
3000
10:02
We can actually figure it out
240
587000
2000
10:04
using really basic signal processing.
241
589000
2000
10:06
Here's a simple way to do it.
242
591000
2000
10:08
Well, a reasonable expectation
243
593000
2000
10:10
is that somebody's fame in a given period of time
244
595000
2000
10:12
should be roughly the average of their fame before
245
597000
2000
10:14
and their fame after.
246
599000
2000
10:16
So that's sort of what we expect.
247
601000
2000
10:18
And we compare that to the fame that we observe.
248
603000
3000
10:21
And we just divide one by the other
249
606000
2000
10:23
to produce something we call a suppression index.
250
608000
2000
10:25
If the suppression index is very, very, very small,
251
610000
3000
10:28
then you very well might be being suppressed.
252
613000
2000
10:30
If it's very large, maybe you're benefiting from propaganda.
253
615000
3000
10:34
JM: Now you can actually look at
254
619000
2000
10:36
the distribution of suppression indexes over whole populations.
255
621000
3000
10:39
So for instance, here --
256
624000
2000
10:41
this suppression index is for 5,000 people
257
626000
2000
10:43
picked in English books where there's no known suppression --
258
628000
2000
10:45
it would be like this, basically tightly centered on one.
259
630000
2000
10:47
What you expect is basically what you observe.
260
632000
2000
10:49
This is distribution as seen in Germany --
261
634000
2000
10:51
very different, it's shifted to the left.
262
636000
2000
10:53
People talked about it twice less as it should have been.
263
638000
3000
10:56
But much more importantly, the distribution is much wider.
264
641000
2000
10:58
There are many people who end up on the far left on this distribution
265
643000
3000
11:01
who are talked about 10 times fewer than they should have been.
266
646000
3000
11:04
But then also many people on the far right
267
649000
2000
11:06
who seem to benefit from propaganda.
268
651000
2000
11:08
This picture is the hallmark of censorship in the book record.
269
653000
3000
11:11
ELA: So culturomics
270
656000
2000
11:13
is what we call this method.
271
658000
2000
11:15
It's kind of like genomics.
272
660000
2000
11:17
Except genomics is a lens on biology
273
662000
2000
11:19
through the window of the sequence of bases in the human genome.
274
664000
3000
11:22
Culturomics is similar.
275
667000
2000
11:24
It's the application of massive-scale data collection analysis
276
669000
3000
11:27
to the study of human culture.
277
672000
2000
11:29
Here, instead of through the lens of a genome,
278
674000
2000
11:31
through the lens of digitized pieces of the historical record.
279
676000
3000
11:34
The great thing about culturomics
280
679000
2000
11:36
is that everyone can do it.
281
681000
2000
11:38
Why can everyone do it?
282
683000
2000
11:40
Everyone can do it because three guys,
283
685000
2000
11:42
Jon Orwant, Matt Gray and Will Brockman over at Google,
284
687000
3000
11:45
saw the prototype of the Ngram Viewer,
285
690000
2000
11:47
and they said, "This is so fun.
286
692000
2000
11:49
We have to make this available for people."
287
694000
3000
11:52
So in two weeks flat -- the two weeks before our paper came out --
288
697000
2000
11:54
they coded up a version of the Ngram Viewer for the general public.
289
699000
3000
11:57
And so you too can type in any word or phrase that you're interested in
290
702000
3000
12:00
and see its n-gram immediately --
291
705000
2000
12:02
also browse examples of all the various books
292
707000
2000
12:04
in which your n-gram appears.
293
709000
2000
12:06
JM: Now this was used over a million times on the first day,
294
711000
2000
12:08
and this is really the best of all the queries.
295
713000
2000
12:10
So people want to be their best, put their best foot forward.
296
715000
3000
12:13
But it turns out in the 18th century, people didn't really care about that at all.
297
718000
3000
12:16
They didn't want to be their best, they wanted to be their beft.
298
721000
3000
12:19
So what happened is, of course, this is just a mistake.
299
724000
3000
12:22
It's not that strove for mediocrity,
300
727000
2000
12:24
it's just that the S used to be written differently, kind of like an F.
301
729000
3000
12:27
Now of course, Google didn't pick this up at the time,
302
732000
3000
12:30
so we reported this in the science article that we wrote.
303
735000
3000
12:33
But it turns out this is just a reminder
304
738000
2000
12:35
that, although this is a lot of fun,
305
740000
2000
12:37
when you interpret these graphs, you have to be very careful,
306
742000
2000
12:39
and you have to adopt the base standards in the sciences.
307
744000
3000
12:42
ELA: People have been using this for all kinds of fun purposes.
308
747000
3000
12:45
(Laughter)
309
750000
7000
12:52
Actually, we're not going to have to talk,
310
757000
2000
12:54
we're just going to show you all the slides and remain silent.
311
759000
3000
12:57
This person was interested in the history of frustration.
312
762000
3000
13:00
There's various types of frustration.
313
765000
3000
13:03
If you stub your toe, that's a one A "argh."
314
768000
3000
13:06
If the planet Earth is annihilated by the Vogons
315
771000
2000
13:08
to make room for an interstellar bypass,
316
773000
2000
13:10
that's an eight A "aaaaaaaargh."
317
775000
2000
13:12
This person studies all the "arghs,"
318
777000
2000
13:14
from one through eight A's.
319
779000
2000
13:16
And it turns out
320
781000
2000
13:18
that the less-frequent "arghs"
321
783000
2000
13:20
are, of course, the ones that correspond to things that are more frustrating --
322
785000
3000
13:23
except, oddly, in the early 80s.
323
788000
3000
13:26
We think that might have something to do with Reagan.
324
791000
2000
13:28
(Laughter)
325
793000
2000
13:30
JM: There are many usages of this data,
326
795000
3000
13:33
but the bottom line is that the historical record is being digitized.
327
798000
3000
13:36
Google has started to digitize 15 million books.
328
801000
2000
13:38
That's 12 percent of all the books that have ever been published.
329
803000
2000
13:40
It's a sizable chunk of human culture.
330
805000
3000
13:43
There's much more in culture: there's manuscripts, there newspapers,
331
808000
3000
13:46
there's things that are not text, like art and paintings.
332
811000
2000
13:48
These all happen to be on our computers,
333
813000
2000
13:50
on computers across the world.
334
815000
2000
13:52
And when that happens, that will transform the way we have
335
817000
3000
13:55
to understand our past, our present and human culture.
336
820000
2000
13:57
Thank you very much.
337
822000
2000
13:59
(Applause)
338
824000
3000
ABOUT THE SPEAKERS
Jean-Baptiste Michel - Data researcherJean-Baptiste Michel looks at how we can use large volumes of data to better understand our world.
Why you should listen
Jean-Baptiste Michel holds joint academic appointments at Harvard (FQEB Fellow) and Google (Visiting Faculty). His research focusses on using large volumes of data as tools that help better understand the world around us -- from the way diseases progress in patients over years, to the way cultures change in human societies over centuries. With his colleague Erez Lieberman Aiden, Jean-Baptiste is a Founding Director of Harvard's Cultural Observatory, where their research team pioneers the use of quantitative methods for the study of human culture, language and history. His research was featured on the covers of Science and Nature, on the front pages of the New York Times and the Boston Globe, in The Economist, Wired and many other venues. The online tool he helped create -- ngrams.googlelabs.com -- was used millions of times to browse cultural trends. Jean-Baptiste is an Engineer from Ecole Polytechnique (Paris), and holds an MS in Applied Mathematics and a PhD in Systems Biology from Harvard.
Jean-Baptiste Michel | Speaker | TED.com
Erez Lieberman Aiden - Researcher
Erez Lieberman Aiden pursues a broad range of research interests, spanning genomics, linguistics, mathematics ...
Why you should listen
Erez Lieberman Aiden is a fellow at the Harvard Society of Fellows and Visiting Faculty at Google. His research spans many disciplines and has won numerous awards, including recognition for one of the top 20 "Biotech Breakthroughs that will Change Medicine", by Popular Mechanics; the Lemelson-MIT prize for the best student inventor at MIT; the American Physical Society's Award for the Best Doctoral Dissertation in Biological Physics; and membership in Technology Review's 2009 TR35, recognizing the top 35 innovators under 35. His last three papers -- two with JB Michel -- have all appeared on the cover of Nature and Science.
Erez Lieberman Aiden | Speaker | TED.com