WEBVTT
Kind: captions
Language: en-GB

1
00:00:00.089 --> 00:00:04.260
Machine translation is incredibly difficult.
And to prove that, I will now read this introduction

2
00:00:04.260 --> 00:00:07.470
again, after it’s been sent through Google’s
translator -- currently one of the best in

3
00:00:07.470 --> 00:00:10.690
the world -- and then translated back into
English.

4
00:00:10.690 --> 00:00:14.990
Machine translation is very difficult. Back
then translated into English - is one of the

5
00:00:14.990 --> 00:00:19.449
best in the world right now - it is to prove
that, after being sent through Google's translator,

6
00:00:19.449 --> 00:00:21.230
I'll read this again introduced.

7
00:00:21.230 --> 00:00:25.949
Okay, I chose a difficult language, but each
one I tried introduced subtle errors in different

8
00:00:25.949 --> 00:00:30.859
ways. Via Chinese, it had been translated
by “Google hair”. Via French, the introduction

9
00:00:30.859 --> 00:00:35.570
became a “he”, not an “it”. And those
sentences were incredibly simple.

10
00:00:35.570 --> 00:00:38.520
Folks who only speak one language -- and I
am embarrassed to say that’s a group that

11
00:00:38.520 --> 00:00:42.079
includes me, I’m sorry -- folks who only
speak one language often assume that you can

12
00:00:42.079 --> 00:00:46.710
open a translation dictionary, pick an appropriate
word, faff around with the grammar a bit,

13
00:00:46.710 --> 00:00:50.890
and have a functional sentence in another
language. For simple sentences, yes, that’s

14
00:00:50.890 --> 00:00:55.840
true: but very few sentences in the real world
are actually that simple.

15
00:00:55.840 --> 00:00:59.980
Google recently released a paper about how
they’d reduced machine translation to a

16
00:00:59.980 --> 00:01:06.679
problem in vector space mathematics, representations
of concepts in an abstract language space.

17
00:01:06.679 --> 00:01:10.829
Which is great for mapping concepts to words,
and it’ll even deal well with homographs,

18
00:01:10.829 --> 00:01:15.560
identical words that mean completely different
things. You can deal with those through context:

19
00:01:15.560 --> 00:01:19.770
the days of “hydraulic ram” being translated
as “water sheep” are pretty much in the

20
00:01:19.770 --> 00:01:20.420
past.

21
00:01:20.420 --> 00:01:22.399
[OFF SCREEN LAUGHTER]

22
00:01:22.399 --> 00:01:24.369
Spot the engineer.

23
00:01:24.369 --> 00:01:28.359
For formal, technical documents, it might
even start to work well.

24
00:01:28.359 --> 00:01:31.929
But for more casual communication, it’s
not so easy.

25
00:01:31.929 --> 00:01:36.219
Heck, translating between British English
and American English isn’t always easy.

26
00:01:36.219 --> 00:01:40.679
Not because your car’s “hood” is our
“bonnet”, but because “that’s a brave

27
00:01:40.679 --> 00:01:44.999
idea” isn’t a compliment in British English,
it means you’re a prat and your idea is

28
00:01:44.999 --> 00:01:45.389
impossible.

29
00:01:45.389 --> 00:01:50.929
There are concepts which don’t quite match
between languages. “Bonne nuit” might

30
00:01:50.929 --> 00:01:55.499
literally mean the same as “buenas noches”
-- I’m sorry about my pronunciation there

31
00:01:55.499 --> 00:02:00.249
-- but one is meant for saying goodnight at
bedtime and the other’s for saying hello

32
00:02:00.249 --> 00:02:03.259
or goodbye at any point after dark.

33
00:02:03.259 --> 00:02:09.069
Then you have the concepts that don’t translate
between languages at all. In French, “you”

34
00:02:09.069 --> 00:02:12.360
translates as “vous” if it’s someone
you should be respectful towards, and “tu”

35
00:02:12.360 --> 00:02:17.900
if it’s a more casual conversation. Or if
you’re talking to God. No, really. God is

36
00:02:17.900 --> 00:02:22.750
“tu”. A computer will crush both of those
to “you” when translating to other languages,

37
00:02:22.750 --> 00:02:27.120
and it won’t have any idea which of them
to use when translating into French.

38
00:02:27.120 --> 00:02:32.599
And that is just a simple “honorifics”
system. Korean has a much more complicated

39
00:02:32.599 --> 00:02:38.189
set of pronouns for all sorts of situations.
Remember this? That repeated line: oppan Gangnam

40
00:02:38.189 --> 00:02:43.310
style. The English translation of “oppa”
is usually “a woman’s older brother”:

41
00:02:43.310 --> 00:02:48.079
but in everyday speech, “oppa” is used
to refer to someone based on a series of complicated

42
00:02:48.079 --> 00:02:53.730
and fuzzy rules that make instinctive sense
to native speakers. To make it worse, PSY

43
00:02:53.730 --> 00:02:58.829
is referring to himself in the third person
there, which sounds really weird when translated

44
00:02:58.829 --> 00:03:05.049
out of Korean. There is no way to translate
all of the meaning in those words into one

45
00:03:05.049 --> 00:03:06.890
English sentence.

46
00:03:06.890 --> 00:03:12.340
Then you have the problem of shared expectations.
English-speaking cultures tend to be monochronic:

47
00:03:12.340 --> 00:03:17.469
if you make an appointment to meet someone
at 11am, you are expected to be there at about

48
00:03:17.469 --> 00:03:22.290
11am. I mean, groups of friends can often
get around this -- “the party starts at

49
00:03:22.290 --> 00:03:26.680
6” often means people will turn up anywhere
from 6:30 to 9. But imagine if that lack of

50
00:03:26.680 --> 00:03:32.329
punctuality, and that acceptance of a lack
of punctuality, expanded to all aspects of

51
00:03:32.329 --> 00:03:38.219
everyday life. Welcome to the rest of the
world. Massive parts of this planet run on

52
00:03:38.219 --> 00:03:44.189
what is called polychronic time. Two appointments
at the same time? That’s fine, they’ll

53
00:03:44.189 --> 00:03:45.670
understand. And they will understand.

54
00:03:45.670 --> 00:03:50.999
Needless to say, there is often quite significant
culture clash when monochronic and polychronic

55
00:03:50.999 --> 00:03:55.489
people meet. But a machine translation isn’t
going to see an English sentence like “I’ll

56
00:03:55.489 --> 00:04:00.319
meet you at 7pm” and add a note for someone
in a polychronic culture that, no, they really

57
00:04:00.319 --> 00:04:03.389
do mean 7pm, and they’re going to be annoyed
if you’re late.

58
00:04:03.389 --> 00:04:07.900
Ultimately, to accurately translate something,
you don’t just need to know how words map

59
00:04:07.900 --> 00:04:12.389
to concepts: you need to understand social
structures, subtext, nuance, innuendo. You

60
00:04:12.389 --> 00:04:18.370
need at least a basic theory of mind: the
idea that the speaker and the listener both

61
00:04:18.370 --> 00:04:23.830
have beliefs and desires expressed by the
particular words they’ve chosen. Translators

62
00:04:23.830 --> 00:04:28.349
need to be able to ask questions of the original
author, so you can check that the subtleties

63
00:04:28.349 --> 00:04:32.410
that you have to add to their work reflect
their intention.

64
00:04:32.410 --> 00:04:36.610
The problem isn’t that language is messy
-- computers can cope with messy, heck, they

65
00:04:36.610 --> 00:04:41.099
can pretty much solve CAPTCHAs better than
humans these days. The problem is that language

66
00:04:41.099 --> 00:04:47.360
relies on intent, on shared secrets, on group
identity, and on hidden knowledge. Machine

67
00:04:47.360 --> 00:04:51.169
translation is a useful tool, don’t get
me wrong, but trying to get a machine to translate

68
00:04:51.169 --> 00:04:54.100
better than a human is… a brave idea.

69
00:04:55.500 --> 00:04:57.500
[Translating thee subtitles? Add your name here!]