1
00:00:00,080 --> 00:00:04,960
Google Translate is an amazing tool but

2
00:00:03,120 --> 00:00:09,719
it also kind of sucks it's a service that's come a long way over the last 18

3
00:00:07,759 --> 00:00:13,240
years but it's still nowhere near the perfect Universal translator we're used

4
00:00:11,719 --> 00:00:18,000
to seeing in science fiction like Star Trek and hitchhiker Guide to the Galaxy

5
00:00:16,000 --> 00:00:23,080
but how far away from that kind of seamless translation technology are we

6
00:00:20,359 --> 00:00:27,400
really Google Translate is an example of machine translation a subfield of

7
00:00:25,519 --> 00:00:32,480
computational linguistics that has been under development since the 1950s early

8
00:00:30,199 --> 00:00:36,480
machine translators used a rule-based approach which required programmers to

9
00:00:34,280 --> 00:00:41,440
explicitly describe every rule governing a language which was both labor

10
00:00:38,360 --> 00:00:43,200
intensive and flawed in the 1980s

11
00:00:41,440 --> 00:00:47,480
statistical translation models became popular these used probability and

12
00:00:45,480 --> 00:00:51,719
heuristics to determine the best translation for a given phrase which was

13
00:00:49,840 --> 00:00:55,559
more efficient but they struggled to translate between languages with

14
00:00:53,399 --> 00:01:00,519
significant differences in terms of grammar and word order JM is a pretty

15
00:00:58,800 --> 00:01:06,720
normal French phrase but its word order isn't I love you it's

16
00:01:03,480 --> 00:01:09,119
I you love this is one of the reasons

17
00:01:06,720 --> 00:01:14,080
why even relatively fast real-time interpretation programs always come with

18
00:01:11,360 --> 00:01:17,960
a significant delay the software needs to hear the full sentence before it

19
00:01:15,720 --> 00:01:22,280
knows how it's supposed to order the words in its translation when it

20
00:01:20,000 --> 00:01:25,880
launched in 2006 Google translate originally used a statistical

21
00:01:24,000 --> 00:01:29,920
translation model and typically translated Languages by converting them

22
00:01:27,720 --> 00:01:34,680
to English then converting them into into the desired target language in late

23
00:01:32,200 --> 00:01:39,479
2016 Google transitioned its translation service to neural machine translation

24
00:01:36,920 --> 00:01:44,399
which uses a neural network to translate phrases and passages holistically

25
00:01:41,920 --> 00:01:49,119
meaning that it can use the larger text as context when translating specific

26
00:01:46,880 --> 00:01:53,960
phrases Google Translate in particular excels because it's so widely used and

27
00:01:51,719 --> 00:02:00,119
has a massive flow of user feedback to help refine its output that being said

28
00:01:57,479 --> 00:02:05,039
experts do not consider any translation programs today to be capable of fully

29
00:02:02,280 --> 00:02:09,879
automated high quality translation as the technology is still too unreliable

30
00:02:07,560 --> 00:02:14,920
for formal purposes like translating works for publication or interpreting a

31
00:02:12,360 --> 00:02:19,480
witness statement during a court hearing however professional translators do use

32
00:02:17,280 --> 00:02:24,680
a fair number of automated Tools in what is called computer assisted translation

33
00:02:22,080 --> 00:02:29,959
their cyborgs to be fair accurate enough for a court hearing is a pretty high bar

34
00:02:27,319 --> 00:02:34,599
so how accurate is Google Translate really well it can vary a lot depending

35
00:02:32,280 --> 00:02:38,800
on the language pairing if you're a unilingual English speaker Google

36
00:02:36,879 --> 00:02:43,159
translate might actually seem to work really really well that's because

37
00:02:40,599 --> 00:02:47,879
English has the most speakers out of any human language with almost 1.5 billion

38
00:02:46,000 --> 00:02:51,480
and half of websites are written in English Google translate has tons of

39
00:02:49,840 --> 00:02:56,360
English language reference material to draw from but the majority of the other

40
00:02:54,040 --> 00:03:01,360
132 languages supported by Google translate are only represented by a

41
00:02:58,640 --> 00:03:07,440
small fraction of users and websites according to a 2019 UCLA study English

42
00:03:04,519 --> 00:03:12,000
to Spanish was around 94% accurate but English to Armenian was only 55%

43
00:03:12,519 --> 00:03:19,879
accurate but isn't this accuracy Gap just going to go away with

44
00:03:17,000 --> 00:03:23,920
time well yes and no this technology will inevitably become more accurate as

45
00:03:22,200 --> 00:03:28,319
time goes on but there are several deeper problems with machine translation

46
00:03:25,920 --> 00:03:32,480
that we have yet to really solve we'll tell you what they are after this

47
00:03:29,799 --> 00:03:36,200
message from our sponsor iix it if your laptop is broken or acting funky forget

48
00:03:34,720 --> 00:03:41,360
going out and buying a whole new one repair it with the help of iix it their

49
00:03:38,840 --> 00:03:44,439
exhaustive selection of Parts like ssds and batteries along with their

50
00:03:42,760 --> 00:03:48,519
comprehensive repair guides means opening up your laptop and fixing it is

51
00:03:46,360 --> 00:03:52,920
easier than ever check out iFix it using the link in the description and give

52
00:03:50,439 --> 00:03:57,560
your busted laptop a new lease on life it's tempting to think that the route to

53
00:03:54,439 --> 00:03:59,439
a universal translator is just more data

54
00:03:57,560 --> 00:04:03,680
and that would definitely help in arm media case but even with our extremely

55
00:04:02,000 --> 00:04:08,239
sophisticated neural network and plenty of training data machine translators

56
00:04:06,000 --> 00:04:13,879
tend to struggle in a few key areas of human communication most notably slang

57
00:04:11,319 --> 00:04:17,359
jokes and figurative language a lot of human communication relies on

58
00:04:15,439 --> 00:04:22,919
abstraction and double meanings especially metaphors and idioms the

59
00:04:19,440 --> 00:04:24,720
French phrase okam twam I'm sorry is a

60
00:04:22,919 --> 00:04:29,240
Whimsical idiom typically used to describe the height of young children

61
00:04:27,040 --> 00:04:33,960
the rough English equivalent is knee high to a grasshopper but if you

62
00:04:31,280 --> 00:04:39,479
translate it literally it means three apples High Google translate interprets

63
00:04:36,960 --> 00:04:45,120
this French phrase as little person which is sort of right but also very

64
00:04:41,840 --> 00:04:47,800
very wrong similarly the humor of it

65
00:04:45,120 --> 00:04:51,680
ain't rocket surgery just isn't going to translate well without a human being

66
00:04:49,880 --> 00:04:56,039
willing to put in the groundwork to find a suitable cultural equivalent because

67
00:04:53,919 --> 00:05:00,400
French people don't use rocket science or brain surgery as benchmarks for

68
00:04:58,120 --> 00:05:04,000
difficulty they use sourc now this isn't an impossible hurdle for

69
00:05:02,120 --> 00:05:08,000
a machine translator to handle because idioms tend to be said the exact same

70
00:05:05,880 --> 00:05:12,240
way every time you could teach the software how to translate a long list of

71
00:05:10,000 --> 00:05:16,720
specific idioms so long as you are willing to put enough resources into it

72
00:05:14,680 --> 00:05:21,919
teaching it to consistently translate word play and jokes however might very

73
00:05:19,759 --> 00:05:25,319
well be impossible sing is always a real headache for translators both machine

74
00:05:23,440 --> 00:05:29,919
and human because it's typically used by relatively small Niche subcultures and

75
00:05:27,960 --> 00:05:35,479
it's unlikely to wind up in Main language repositories like dictionaries

76
00:05:32,199 --> 00:05:37,759
and thorth slang also tends to change

77
00:05:35,479 --> 00:05:41,800
quickly and rely on community specific cultural references if you were to try

78
00:05:39,919 --> 00:05:45,919
and translate a slang heavy Kendrick Lamar song by pushing it through a

79
00:05:43,560 --> 00:05:51,000
machine translator the result would not be just aesthetically questionable but

80
00:05:48,600 --> 00:05:55,120
also basically incoherent to a person unfamiliar with American rap culture

81
00:05:53,440 --> 00:05:58,880
helping a machine translator to understand slang is again possible it

82
00:05:57,400 --> 00:06:03,280
would just be a bit expensive and require near constant updating where

83
00:06:01,520 --> 00:06:07,800
things get tricky is that there's a slight difference between a translation

84
00:06:05,039 --> 00:06:12,720
being accurate and a translation being good especially when it comes to Art

85
00:06:10,599 --> 00:06:16,960
machine translation can often be technically correct but still fail to

86
00:06:14,680 --> 00:06:22,560
communicate the tone and cultural connotations of the original on a purely

87
00:06:19,960 --> 00:06:27,319
technical level how are you what's up and how's it hanging all mean basically

88
00:06:24,919 --> 00:06:31,000
the same thing but one is a double on Tandra and a weird thing to say to your

89
00:06:28,880 --> 00:06:34,919
grandma we might be happy enough with a machine translation of a menu if it just

90
00:06:33,280 --> 00:06:39,680
accurately communicates what food is available but a translated novel or poem

91
00:06:38,000 --> 00:06:43,880
needs to find a good balance between accuracy and Aesthetics in order to

92
00:06:41,840 --> 00:06:48,479
create a similar experience as the original this means that a literary

93
00:06:45,800 --> 00:06:52,919
translator needs to be creative and make astute artistic judgments in the same

94
00:06:50,919 --> 00:06:56,840
way that an author does the basic problem with translation software is

95
00:06:54,560 --> 00:07:01,120
that there's no mind behind it that truly understands the purpose or intent

96
00:06:58,919 --> 00:07:04,720
of the words its processing an optimistic goal of machine translation

97
00:07:02,879 --> 00:07:08,160
as a discipline is to one day have translation software that is so

98
00:07:06,080 --> 00:07:12,800
sophisticated and nuanced that it only needs a human Editor to check its work

99
00:07:10,000 --> 00:07:17,160
and fix whatever errors they find more pessimistically highquality fully

100
00:07:14,800 --> 00:07:22,000
Automated machine translation might require the development of something

101
00:07:18,879 --> 00:07:23,560
like AGI artificial general intelligence

102
00:07:22,000 --> 00:07:28,800
a machine capable of human-like intelligence and making Nuance judgment

103
00:07:25,479 --> 00:07:31,800
calls so Google translate might need to

104
00:07:28,800 --> 00:07:33,960
be capable of forming opinions before

105
00:07:31,800 --> 00:07:37,960
it's ready to translate poetry thanks for watching guys if you like this video

106
00:07:35,680 --> 00:07:43,240
why don't you check out our one on X's or Twitter's Community notes program

107
00:07:40,520 --> 00:07:46,360
it's actually pretty cool and maybe the future of the internet