WEBVTT Kind: captions; Language: fi

1
00:00:05.200 --> 00:00:10.610
we can get started to locate while thank you so much to the language

2
00:00:10.610 --> 00:00:16.330
campus here at the University of Jyväskylä for the opportunity to speak to day

3
00:00:16.330 --> 00:00:21.970
about de mystifying machine translation and improving machine translation literacy

4
00:00:21.970 --> 00:00:26.080
my name is Lynn Bowker and my regular job is as

5
00:00:26.080 --> 00:00:31.380
professor at the school of translation on interpretation the at the university of ottawa

6
00:00:31.380 --> 00:00:36.380
in canada but for the months of november and december i'm being host

7
00:00:36.380 --> 00:00:41.540
of as visiting fellow here at the university of Jyväskylä

8
00:00:41.540 --> 00:00:47.690
so i'm going to divide to day's talk into two main parts

9
00:00:47.690 --> 00:00:52.310
the first thing that i'm going to do is little bit of crash course

10
00:00:52.310 --> 00:00:57.150
to de mystify machine translation and talk little bit about

11
00:00:57.150 --> 00:01:01.750
what is nerle machine translation or how to day driven approaches to machine

12
00:01:01.750 --> 00:01:06.910
translation work ah what is all the rhythmic bias and little bit of human computer

13
00:01:06.910 --> 00:01:11.860
interaction and essentially this is sort of crash course

14
00:01:11.860 --> 00:01:17.330
inrushing translation literacy so thought did at first and then afterwards

15
00:01:17.330 --> 00:01:22.110
out when you understand what coarse and wishing physician literacy might look like portly

16
00:01:22.110 --> 00:01:26.790
stood in o'er sort of bud trimmed down version ah then we can

17
00:01:26.790 --> 00:01:31.790
talk more about how you might like to participate in the teaching of machine translation

18
00:01:31.790 --> 00:01:36.880
received two others whether their your students or whether you work with other groups that might benefit

19
00:01:36.880 --> 00:01:41.670
from machine translation literacy instruction like tiens or

20
00:01:41.670 --> 00:01:46.540
new comers to country or other people that he might encounter

21
00:01:46.540 --> 00:01:52.040
so those will be the too kind of hops of the talk and da we can take questions

22
00:01:52.040 --> 00:01:57.310
after each half of that sort of make sense so that you have questions up popped them in the chat

23
00:01:57.310 --> 00:02:01.790
until we can adjust them at the end of it half of the talk

24
00:02:01.790 --> 00:02:03.630
k.

25
00:02:03.630 --> 00:02:08.330
so for de mystifying machine translation.

26
00:02:08.330 --> 00:02:13.440
ah me when we talk about machine translation where really talking about that

27
00:02:13.440 --> 00:02:18.750
category or type of tool that automatically translates

28
00:02:18.750 --> 00:02:24.060
text from one language such as french into another language such as english

29
00:02:24.060 --> 00:02:28.530
and it's little bit up to you whether you ask thee tool to translate just

30
00:02:28.530 --> 00:02:33.680
word for you whole sentenced may be longer taxed but the point is that

31
00:02:33.680 --> 00:02:38.570
whatever tax you offer to the machine it will translate the whole

32
00:02:38.570 --> 00:02:43.630
thing for you with really know intervention on your part so in that

33
00:02:43.630 --> 00:02:48.700
way it's different from other types of tools that you might have used alike electronic

34
00:02:48.700 --> 00:02:53.520
dictionaries or concord answers or translation memory tools those are all kind of

35
00:02:53.520 --> 00:02:58.520
morn the category of computer assisted translation they do some of the work for you

36
00:02:58.520 --> 00:03:03.710
but there's lot of interaction out between you and the tool machine translation

37
00:03:03.710 --> 00:03:09.110
there can reason interaction usually either before or after the translation process

38
00:03:09.110 --> 00:03:13.600
but the translation process itself is carried out automatically so of

39
00:03:13.600 --> 00:03:18.810
course one of the best known examples of the machine translation tool is google translate is one that probably

40
00:03:18.810 --> 00:03:23.520
almost everyone has heard of so that's the type of tool that were talking about

41
00:03:23.520 --> 00:03:29.010
you might wonder why we talk about machine translation

42
00:03:29.010 --> 00:03:33.630
and not computer translation and this is really just kind of

43
00:03:33.630 --> 00:03:38.570
foss throwback term machine translation technology is much

44
00:03:38.570 --> 00:03:44.060
older than people might realize we've sort of and countered it in our daily lives

45
00:03:44.060 --> 00:03:48.620
only in the past ten to fifteen years since tools like google translate

46
00:03:48.620 --> 00:03:53.620
became free online tools but the technology itself goes back much

47
00:03:53.620 --> 00:03:58.680
further back to the post second world war period ah that's when

48
00:03:58.680 --> 00:04:03.830
computers were just emerging and when people talked about odd

49
00:04:03.830 --> 00:04:08.630
developing applications machine translation was ironically one of the first things that

50
00:04:08.630 --> 00:04:13.530
they tried to do so that at that time you can see that confused her

51
00:04:13.530 --> 00:04:18.710
swear really machines they were gigantic they filled an entire room they had all

52
00:04:18.710 --> 00:04:23.700
kinds of tubes and wires and so at that time we talked about computing

53
00:04:23.700 --> 00:04:28.810
machines and so machine translation is as term really just

54
00:04:28.810 --> 00:04:34.030
kind of throwback to that period even though now of course when we talk about

55
00:04:34.030 --> 00:04:39.520
computing machines we talk about computers but just to make that distinction that am

56
00:04:39.520 --> 00:04:44.040
opened with that machine translation is fully automatic and computer aided

57
00:04:44.040 --> 00:04:49.730
translation is other types of tools that are used by translators but

58
00:04:49.730 --> 00:04:54.130
that require assistance from translators to be used effective

59
00:04:54.130 --> 00:04:55.660
k.

60
00:04:55.660 --> 00:05:00.520
another really important thing to realize what is that machine

61
00:05:00.520 --> 00:05:06.170
translation is category of tools it's not single tool

62
00:05:06.170 --> 00:05:10.060
of course google translate is very well known ah it's very

63
00:05:10.060 --> 00:05:14.950
widely used recent estimate was that their over one billion users of

64
00:05:14.950 --> 00:05:19.620
machine of cool translate machine translation jewel on but of course

65
00:05:19.620 --> 00:05:24.710
it's not the only one there are whole host of others neither

66
00:05:24.710 --> 00:05:29.540
are it is also not comprehensive list but just some examples of other

67
00:05:29.540 --> 00:05:34.890
tools that are in the same category so all of these are machine translation tools

68
00:05:34.890 --> 00:05:39.010
and this is sometimes are worth keeping in mind

69
00:05:39.010 --> 00:05:43.520
because even if our first instinct is to go to gobble

70
00:05:43.520 --> 00:05:48.550
we should realize that just like any other type of product different tools have

71
00:05:48.550 --> 00:05:53.900
different strengths souf we think of other types of products like cars in perhaps

72
00:05:53.900 --> 00:05:58.890
you are fan of nice on or of forward or of

73
00:05:58.890 --> 00:06:03.640
in another brand of car because it has certain features that you like while machine translation is

74
00:06:03.640 --> 00:06:08.930
little bit similar you might find that different tools perform better or worse

75
00:06:08.930 --> 00:06:13.880
depending on the task that you have in mind or the language combination that you were quiff

76
00:06:13.880 --> 00:06:18.830
wore the subject field that your translating so

77
00:06:18.830 --> 00:06:23.930
all of the tools that i'd listed here are free and on wind and your some

78
00:06:23.930 --> 00:06:29.130
other different characteristic so good goal certainly has the widest range of languages

79
00:06:29.130 --> 00:06:33.790
available including quite few last widely used languages

80
00:06:33.790 --> 00:06:38.630
so if you are someone working with language um pet pair that is

81
00:06:38.630 --> 00:06:43.660
not very widely used you might have better chance of finding of that pair handled by google

82
00:06:43.660 --> 00:06:48.250
than by some other tools ah being my christophe translator

83
00:06:48.250 --> 00:06:53.580
is an example that like to point out because it offers different varieties

84
00:06:53.580 --> 00:06:57.850
of some languages so in terms of french for example

85
00:06:57.850 --> 00:07:02.830
you could actually choose to translate into the canadian version of french

86
00:07:02.830 --> 00:07:07.210
or other versions of french so that's kind of um nice option for people

87
00:07:07.210 --> 00:07:11.900
who are users of particular variety of language

88
00:07:11.900 --> 00:07:17.030
deep isle is another tool that is becoming very popular and has possibly

89
00:07:17.030 --> 00:07:21.950
got the highest quality for many western european languages it doesn't do as

90
00:07:21.950 --> 00:07:26.740
many languages as bugle but the once that it does it does quite while so on the whole

91
00:07:26.740 --> 00:07:31.460
deep is known as being high quality tool on papago is

92
00:07:31.460 --> 00:07:36.670
tool that strong for language like korean ca by do for chinese

93
00:07:36.670 --> 00:07:41.260
at yon decks tends to be very strong for eastern european languages so again it

94
00:07:41.260 --> 00:07:46.380
make sense to pick your machine translation tool based on your immediate needs

95
00:07:46.380 --> 00:07:51.210
and to realize that while google might do the job its also not the only

96
00:07:51.210 --> 00:07:56.020
auction out there because think sometimes we forget to look beyond the

97
00:07:56.020 --> 00:08:00.770
obvious has so there are whole host of jewels out there to choose from

98
00:08:00.770 --> 00:08:06.340
saw my adviseth ah with regard to transition tools

99
00:08:06.340 --> 00:08:11.140
is that to out in why not try more than one if you're not happy with the results

100
00:08:11.140 --> 00:08:15.860
that you're getting from particular tool whether it's kugler another tool in

101
00:08:15.860 --> 00:08:20.810
of they are free their easily available so why not try another and

102
00:08:20.810 --> 00:08:25.950
you might be happier with the results that you get from another tool and its also important

103
00:08:25.950 --> 00:08:30.860
to remember that these tools are under active development so they

104
00:08:30.860 --> 00:08:35.850
are changing all the time if your disappointed with the results that you get from

105
00:08:35.850 --> 00:08:40.790
tool on particular occasion don't write it off completely and say why never

106
00:08:40.790 --> 00:08:45.970
guinea used not to again it it did such disastrous job may be the text that you're working

107
00:08:45.970 --> 00:08:50.790
with wasn't just well suited to the language combination offered by the tool

108
00:08:50.790 --> 00:08:56.020
order or the i'm no kind of text type of but each translation

109
00:08:56.020 --> 00:09:00.420
jobs different so don't completely give up on tool because you have one bad

110
00:09:00.420 --> 00:09:05.150
experience so little bit of trial and ere think will help you

111
00:09:05.150 --> 00:09:09.730
figure out how which tool works best for you the tides attacks that you normally work

112
00:09:09.730 --> 00:09:14.540
with and just like any other type of tool om

113
00:09:14.540 --> 00:09:19.350
intellectual it's good for one job might not be the right tool for the next on so get

114
00:09:19.350 --> 00:09:23.960
into the habit of really evaluating every time you have translation task to do

115
00:09:23.960 --> 00:09:28.590
in which till will be best suited for the task at hand

116
00:09:28.590 --> 00:09:31.970
think.

117
00:09:31.970 --> 00:09:37.202
so it kind of raises the question of little bit how does machine translation work

118
00:09:37.202 --> 00:09:42.086
and why don't we get the same results from all tools if i'm here

119
00:09:42.086 --> 00:09:47.318
recommending to you that you should try different tools with the same text that is

120
00:09:47.318 --> 00:09:52.202
because you'll get different results with each told that you try and why is

121
00:09:52.202 --> 00:09:57.085
that what we get different results are from different machine translation systems and the

122
00:09:57.085 --> 00:10:02.317
reason is sort of because how these tools worked to day so in the past

123
00:10:02.317 --> 00:10:07.201
when talked about cha <unk> enow just after the second world war when

124
00:10:07.201 --> 00:10:12.433
we were just starting to develop machine translation systems the approach that was taken at

125
00:10:12.433 --> 00:10:17.317
that time was very much based on linguistics thee developers of the tools thought

126
00:10:17.317 --> 00:10:22.200
that if they could build big enough dictionary and if they could and.

127
00:10:22.200 --> 00:10:27.140
hold grammatical rules about how to put different words together

128
00:10:27.140 --> 00:10:31.360
then they would be able to solve the problem of machine translation

129
00:10:31.360 --> 00:10:36.590
so they put their efforts into that about what was called rule based approach where they

130
00:10:36.590 --> 00:10:41.540
were and coating painstakingly all the grammar rules for all the different languages and all

131
00:10:41.540 --> 00:10:46.750
the different combinations of languages and building gigantic dictionaries and

132
00:10:46.750 --> 00:10:51.510
they had limited success but it was definitely limited

133
00:10:51.510 --> 00:10:56.580
we look back now at the quality produced by the early machine transition systems

134
00:10:56.580 --> 00:11:02.260
and we sort of thing while that that sort of disappointing it wasn't that good dame

135
00:11:02.260 --> 00:11:06.600
and its ironic as said that they chose to start with machine

136
00:11:06.600 --> 00:11:11.710
translation as one of the earliest applications of computers because translation is actually very

137
00:11:11.710 --> 00:11:16.820
difficult task on and so that it was

138
00:11:16.820 --> 00:11:21.520
<unk> of strange decision to sort of go for one of the hardest tasks rated the beginning

139
00:11:21.520 --> 00:11:26.610
but eventually what they realized was has perhaps there's

140
00:11:26.610 --> 00:11:31.650
better way instead of trying to make computers process

141
00:11:31.650 --> 00:11:37.180
language in way that similar to how people process language maybe we could

142
00:11:37.180 --> 00:11:41.570
ah let the computers do what they're good at instead and so what we

143
00:11:41.570 --> 00:11:46.890
see now is really sharp turned towards him since about the year

144
00:11:46.890 --> 00:11:51.890
now about to turn of the millennium we've seen ah completely different approach

145
00:11:51.890 --> 00:11:57.030
to machine translation instead of looking at linguistic are rules

146
00:11:57.030 --> 00:12:02.200
and and to repositories of of lexicon

147
00:12:02.200 --> 00:12:07.030
we are in an era now where were taking and dated driven approach

148
00:12:07.030 --> 00:12:11.630
to machine translation and it's completely different away

149
00:12:11.630 --> 00:12:18.660
to approach the problem than the real based approach word than how people actually process language

150
00:12:18.660 --> 00:12:24.680
so let's ah look little bit about why the early systems didn't work so well and this is the slide

151
00:12:24.680 --> 00:12:30.340
that my sons as disturbing but hope i'll be able to explain it to you

152
00:12:30.340 --> 00:12:36.240
if we are comparing the way that language processing takes place in people versus computers

153
00:12:36.240 --> 00:12:40.590
one of the really important things that we may to realize is that

154
00:12:40.590 --> 00:12:45.500
people don't just use grammar and dictionaries we actually use all

155
00:12:45.500 --> 00:12:50.570
lot of our real world knowledge to interpret the taxed

156
00:12:50.570 --> 00:12:55.950
out we use context and so here is an essential example

157
00:12:55.950 --> 00:13:00.770
come in french the word awful ca can have to meanings

158
00:13:00.770 --> 00:13:05.770
so it can be translated from french into english either as

159
00:13:05.770 --> 00:13:10.850
lawyer or as avocado so those are to

160
00:13:10.850 --> 00:13:16.030
potential possible and entirely legitimate translations of the word of orca

161
00:13:16.030 --> 00:13:20.650
we can translate it as lawyer or as of cardot so

162
00:13:20.650 --> 00:13:25.750
oh as person have your probably not going to be on

163
00:13:25.750 --> 00:13:30.920
very ah find it very difficult to make the choice right the context will will tell

164
00:13:30.920 --> 00:13:36.050
you pretty clearly weathered the the that correct translation as of cot oak

165
00:13:36.050 --> 00:13:41.090
or whether it's lawyer and we do this by having little bit of real world knowledge like

166
00:13:41.090 --> 00:13:46.070
lawyers are people and we don't eat people or avocado as our

167
00:13:46.070 --> 00:13:50.760
inanimate i'm kind of food and they can't go to

168
00:13:50.760 --> 00:13:55.760
court in make an argument so in we actually taking on will not even

169
00:13:55.760 --> 00:14:01.060
mill second probably to process that information just because we know how the world works

170
00:14:01.060 --> 00:14:06.000
but if we give the same taxed to machine translation system like

171
00:14:06.000 --> 00:14:10.770
monterey animal ca we might naturally gets translation such

172
00:14:10.770 --> 00:14:15.900
as ate lawyer because lawyers in the dictionary and

173
00:14:15.900 --> 00:14:21.160
it doesn't have any real world knowledge so it really um

174
00:14:21.160 --> 00:14:26.180
in kind of hit wall this clothed with call this mountain barrier this rule based approached

175
00:14:26.180 --> 00:14:31.200
machine translation and so now we actually have turned to the whole process

176
00:14:31.200 --> 00:14:35.840
upside down and said computers are not good at real

177
00:14:35.840 --> 00:14:40.840
world knowledge they don't have any and so what can we ask them to do that they

178
00:14:40.840 --> 00:14:46.180
might be good at instead so some of the greatest strengths of computers

179
00:14:46.180 --> 00:14:52.060
ha are not real world knowledge but things like number crunching

180
00:14:52.060 --> 00:14:55.820
and pattern matching and these are actually super powers of

181
00:14:55.820 --> 00:15:00.790
computer stakin process so many numbers so quickly would make

182
00:15:00.790 --> 00:15:06.440
our had spent and they are relentless and finding every instance

183
00:15:06.440 --> 00:15:11.420
of the pattern if you give it gigantic corpus as billions of words than you ask

184
00:15:11.420 --> 00:15:16.760
it to find all the examples of particular word it can do that very quickly in the blink

185
00:15:16.760 --> 00:15:21.820
of been whereas it would take you or years to read through billions of

186
00:15:21.820 --> 00:15:27.100
words of text so we definitely have some strains but computers have their

187
00:15:27.100 --> 00:15:32.370
strength to the point as their very different so we can't approached translation the same way

188
00:15:32.370 --> 00:15:37.410
with computer as we would tackle that task ourselves

189
00:15:37.410 --> 00:15:43.090
so the very hot most recent shank current

190
00:15:43.090 --> 00:15:47.540
approached machine translation it is what we call and data

191
00:15:47.540 --> 00:15:52.780
driven approach so it's playing to those straits of computers its allowing the computers

192
00:15:52.780 --> 00:15:57.620
to have access to enter process large amounts of data because that's

193
00:15:57.620 --> 00:16:02.560
one of the things that computers really good ah so the most current approaches known as

194
00:16:02.560 --> 00:16:07.730
nerle machine translation and uses artificial intelligence

195
00:16:07.730 --> 00:16:12.600
techniques so it has artificial ner on that works against the little diagram here

196
00:16:12.600 --> 00:16:17.890
of narrow network isle which is an network that take them in put

197
00:16:17.890 --> 00:16:23.000
so the light green notes on the left are the in what then in the middle

198
00:16:23.000 --> 00:16:27.590
in in light blue it shows that the computers doing some kind of process

199
00:16:27.590 --> 00:16:32.620
think and it will output something our which shone in yellow on the on the

200
00:16:32.620 --> 00:16:38.240
other side now on this particular example of unknown at work has just one processing

201
00:16:38.240 --> 00:16:43.560
lair in the middle of unknown at work in have many many layers and

202
00:16:43.560 --> 00:16:47.850
it's little bit of black box what happens in there is some kind of mathematical

203
00:16:47.850 --> 00:16:52.750
stuff and it's not essential for us to understand

204
00:16:52.750 --> 00:16:58.250
exactly what's happening just inside the nerle network in order for us to appreciate

205
00:16:58.250 --> 00:17:02.890
some of the more general ways in which machine translation works so don't worry if you don't

206
00:17:02.890 --> 00:17:08.240
understand all of the kind of glory details of nerle network it's enough to realize

207
00:17:08.240 --> 00:17:12.880
that it is indeed driven approach and that it takes in put data in the form of

208
00:17:12.880 --> 00:17:17.840
training data and it will produce sort of proposed translation for you at the

209
00:17:17.840 --> 00:17:23.340
and of the task now the thing that is really important though to understand

210
00:17:23.340 --> 00:17:28.460
is that as input data on the left the

211
00:17:28.460 --> 00:17:33.470
table driven machine translation systems don't just made one or two examples

212
00:17:33.470 --> 00:17:37.930
and they don't as need like hundred or thousand examples in order

213
00:17:37.930 --> 00:17:42.960
to work they really need millions and millions of examples so

214
00:17:42.960 --> 00:17:48.290
were talking about gigantic corpora and at the bottom of the screen

215
00:17:48.290 --> 00:17:53.020
you can see ah sort of ah just ahead of very small

216
00:17:53.020 --> 00:17:58.060
example of law of parallel corpus so it's corpus where we

217
00:17:58.060 --> 00:18:03.180
have text in one language and its translation into another language and

218
00:18:03.180 --> 00:18:08.040
these are aligned soul usually added at the sentence level alms

219
00:18:08.040 --> 00:18:13.510
so that its clear example for the machine that if you encounter text like this

220
00:18:13.510 --> 00:18:18.110
sentence it might look like that in another language so

221
00:18:18.110 --> 00:18:22.970
if we are giving millions and millions of examples to machine

222
00:18:22.970 --> 00:18:28.660
transition system it can learn from them it's sort of artificial learning that it can learn from them

223
00:18:28.660 --> 00:18:33.430
and it can learn what seems to be likely about were reasonable

224
00:18:33.430 --> 00:18:37.880
translation based on all of the examples that essie

225
00:18:37.880 --> 00:18:42.905
ah the challenge with dated driven approaches though is that these approaches are

226
00:18:42.905 --> 00:18:47.930
also sensitive to data and that means that on the one hand yes

227
00:18:47.930 --> 00:18:53.342
we need many many examples but that's not enough we can't just take lots

228
00:18:53.342 --> 00:18:57.980
and lots of taxed growled whatever we find in just throw it

229
00:18:57.980 --> 00:19:03.392
into big pot winnie to put little bit of care into the

230
00:19:03.392 --> 00:19:08.417
choice of the texts because in addition to meeting lots of data because

231
00:19:08.417 --> 00:19:13.442
we meet the right kind if data and in translation this could mean

232
00:19:13.442 --> 00:19:18.080
that you need the right languages obviously if you want to train

233
00:19:18.080 --> 00:19:23.492
machine translation system to translate from finish into english while it doesn't make

234
00:19:23.492 --> 00:19:28.130
lot of sense to to give it information in arabic and.

235
00:19:28.130 --> 00:19:32.710
spanish rates so you have to of course have the right languages but

236
00:19:32.710 --> 00:19:37.790
even more than that you might like to think about the right having the right

237
00:19:37.790 --> 00:19:42.250
subject matter content if you want your system to translate

238
00:19:42.250 --> 00:19:46.220
medical tax than dozen make lot of sense to to train

239
00:19:46.220 --> 00:19:50.630
it with legal texts or some other domain

240
00:19:50.630 --> 00:19:55.860
and so what where finding in this dated driven world of the stated driven

241
00:19:55.860 --> 00:20:00.980
approached machine translation is that finding the right kind of taita

242
00:20:00.980 --> 00:20:06.200
can be bit of challenge and its led us to situation that gets

243
00:20:06.200 --> 00:20:11.430
described as kind of high resorts languages and laura

244
00:20:11.430 --> 00:20:16.850
source language its so high resorts languages ha could mean languages

245
00:20:16.850 --> 00:20:21.710
that are very widely spoken so english is language that this is very widely spoken

246
00:20:21.710 --> 00:20:26.790
around the world so it's not that hard to find data in english

247
00:20:26.790 --> 00:20:32.010
ah if we wanted talk about language pears we might say well english and french is

248
00:20:32.010 --> 00:20:36.870
high resorts language pair because individually both english and french

249
00:20:36.870 --> 00:20:42.390
are widely spoken and there are many contexts where tax or translated

250
00:20:42.390 --> 00:20:47.450
between those languages we think of canada which has english and french as official languages

251
00:20:47.450 --> 00:20:52.150
i'll within the european union english and french both are official language is so

252
00:20:52.150 --> 00:20:57.070
were not challenge to much when it comes to finding resources that

253
00:20:57.070 --> 00:21:01.980
we could use as examples to feed these machine transition systems

254
00:21:01.980 --> 00:21:07.350
when we get into languages that our last widely used so i'm here in finland

255
00:21:07.350 --> 00:21:12.120
at the moment finishes language that is the national language of finland but it's not

256
00:21:12.120 --> 00:21:17.710
widely spoken anywhere else in the world so it's not ah high resorts language

257
00:21:17.710 --> 00:21:22.140
and if we were to think about translating between finished

258
00:21:22.140 --> 00:21:27.350
and another language that's not high highly use them we have

259
00:21:27.350 --> 00:21:32.070
not only to individual languages that her are less widely used but

260
00:21:32.070 --> 00:21:37.300
the translations between them might be difficult to obtain swelled so may be like finishing

261
00:21:37.300 --> 00:21:42.440
greek maybe there's not lot of ah examples available readily available

262
00:21:42.440 --> 00:21:47.340
between those two rank'd so that means that the machine

263
00:21:47.340 --> 00:21:52.090
transition systems for the high resorts languages tend to produce better quality be

264
00:21:52.090 --> 00:21:57.419
because they have more examples to look at and the machine translation between lang which

265
00:21:57.419 --> 00:22:02.392
pierced that are less widely used might have fewer examples to draw from and

266
00:22:02.392 --> 00:22:07.366
then am not reproducing quite as high quality it get even more complicated

267
00:22:07.366 --> 00:22:12.339
than when we won translate very specialize domains because the same thing happens

268
00:22:12.339 --> 00:22:17.313
some domains are very common or we find translation between them soul topics that

269
00:22:17.313 --> 00:22:22.286
are widely talked about in governments tend to have allowed of corpus resources

270
00:22:22.286 --> 00:22:27.260
available so we might see you know in the european union or again the

271
00:22:27.260 --> 00:22:32.233
canadian government whatever the governments are talking about we have examples of what the

272
00:22:32.233 --> 00:22:37.207
government's don't talk about everything where they have their own particular variety of language

273
00:22:37.207 --> 00:22:42.180
to some of your och bureaucratic administrative out language and it may be left.

274
00:22:42.180 --> 00:22:47.840
asked easy to find very technical texts for example alms to put in our corpora

275
00:22:47.840 --> 00:22:52.870
soul when when we're dealing with combination of potentially less widely

276
00:22:52.870 --> 00:22:57.930
used languages and less common domains all of sudden it's quite challenging to come up

277
00:22:57.930 --> 00:23:03.180
with billions of examples and so this has as created situation

278
00:23:03.180 --> 00:23:08.200
where machine translation works better for some languages and topics than for others

279
00:23:08.200 --> 00:23:13.580
something else that we need to keep in mind is this concept

280
00:23:13.580 --> 00:23:18.470
of ah alder rhythmic bias so we might say that

281
00:23:18.470 --> 00:23:23.630
an earl machine translation system is little but like baby when it starts out

282
00:23:23.630 --> 00:23:28.840
like baby it doesn't know anything and what what it's going to learn is

283
00:23:28.840 --> 00:23:33.700
what we teach it so we have to be careful about what we teach it right

284
00:23:33.700 --> 00:23:38.720
just like we have to maybe be careful like we don't swear in front of her children or

285
00:23:38.720 --> 00:23:43.680
alyona we don't to talk about certain topics in front of her children we have to

286
00:23:43.680 --> 00:23:48.660
think about the machine translation system as being assistant it's going to learn

287
00:23:48.660 --> 00:23:54.390
what we teach it so if the training data is not well chosen

288
00:23:54.390 --> 00:23:58.770
if it contains language that we might consider to be sexes store races

289
00:23:58.770 --> 00:24:04.000
then the machine translation system can potentially learned this and replicate

290
00:24:04.000 --> 00:24:08.930
it so i've got some examples and maybe you can tamiya

291
00:24:08.930 --> 00:24:14.060
for if the translation is is correct but am understand that unfinished

292
00:24:14.060 --> 00:24:19.360
theirs as part of gender neutral pronoun so the the pronoun um doesn't

293
00:24:19.360 --> 00:24:24.530
take eye he or she the way that we would have to make that decision in english

294
00:24:24.530 --> 00:24:28.870
so when translating from finished into english the system hoss to

295
00:24:28.870 --> 00:24:34.280
make choice so in day driven system it's going to make choice

296
00:24:34.280 --> 00:24:38.870
by what it usually sees what is more typical what hasn't learned

297
00:24:38.870 --> 00:24:44.280
from the corpus that it has been given and we can see here that it's making very

298
00:24:44.280 --> 00:24:48.840
different choices between he and she depending on what comes now

299
00:24:48.840 --> 00:24:54.150
axed so he is leader hum but she has grandchild

300
00:24:54.150 --> 00:24:59.160
hut he works and she's taking care of the child so it's

301
00:24:59.160 --> 00:25:03.940
turning out language that we you know that its there's no reason that it

302
00:25:03.940 --> 00:25:09.050
should have chosen here she except that thoughts the examples that

303
00:25:09.050 --> 00:25:14.090
it has seen so om if it's problem which

304
00:25:14.090 --> 00:25:19.570
is called elder rhythmic bias or machine bias but the problem is not really

305
00:25:19.570 --> 00:25:24.210
with the tool or the technology its with the data

306
00:25:24.210 --> 00:25:29.460
and the data has been created by people so people are the problem

307
00:25:29.460 --> 00:25:34.340
in the sooner he ho out we are society may be

308
00:25:34.340 --> 00:25:39.150
that doesn't reflected i'd be ideals of abaft having you know

309
00:25:39.150 --> 00:25:44.890
old sexes i'm no races and and because our texts reflect our society

310
00:25:44.890 --> 00:25:50.200
the machine learns from the text that we provide it and so it may thickly means that we need to work

311
00:25:50.200 --> 00:25:55.510
on ourselves the and work on our own on social issues

312
00:25:55.510 --> 00:26:02.230
out but it it underscores the point that it's important to pick the training data carefully

313
00:26:02.230 --> 00:26:06.340
on because the machine will learn what its taught

314
00:26:06.340 --> 00:26:09.630
king.

315
00:26:09.630 --> 00:26:14.870
something else that think we need to ah

316
00:26:14.870 --> 00:26:19.930
be aware of when were using machine translation is but one of our main jobs

317
00:26:19.930 --> 00:26:25.000
is to conduct little bit of risk assessment for when we

318
00:26:25.000 --> 00:26:29.850
decided to use machine transition tool or not we should take

319
00:26:29.850 --> 00:26:34.800
into account the situation of that were in and the reason that were using

320
00:26:34.800 --> 00:26:39.800
the tool because some situations are more or what we might consider high

321
00:26:39.800 --> 00:26:44.890
stakes than others meaning the consequences of getting it wrong could be

322
00:26:44.890 --> 00:26:50.270
quite serious in some cases where as in other contexts is really not to be deal

323
00:26:50.270 --> 00:26:55.070
if the translation isn't so good so we can think of context

324
00:26:55.070 --> 00:26:59.860
such as no visiting doctor for health problem for thee

325
00:26:59.860 --> 00:27:05.660
stakes of having about translation in the context of health care are more serious

326
00:27:05.660 --> 00:27:10.050
than in context such as no translating up

327
00:27:10.050 --> 00:27:15.110
manga comic for entertainment rain if the track desolation of your maga comic

328
00:27:15.110 --> 00:27:20.090
is terrible he might be disappointed but you're not going to you'll suffer

329
00:27:20.090 --> 00:27:25.130
any serious consequences from not so this is something that think

330
00:27:25.130 --> 00:27:30.370
is sa its again it's not really about the tool itself but it's about

331
00:27:30.370 --> 00:27:35.160
our judgment it's about learning to evaluate what's the good

332
00:27:35.160 --> 00:27:40.500
context for using machine translation and what less good context nowt

333
00:27:40.500 --> 00:27:45.520
sometimes of course we also are face to the situation where

334
00:27:45.520 --> 00:27:50.180
well is potentially poor or potentially lesson

335
00:27:50.180 --> 00:27:55.293
perfect translations still better than nor translation at all and that's where we get

336
00:27:55.293 --> 00:28:00.406
into making very difficult position so sometimes we would say let's give machine

337
00:28:00.406 --> 00:28:05.519
translation of try because the alternative is no translation and that certainly not helpful

338
00:28:05.519 --> 00:28:10.632
out all om song in come in and in second to odd

339
00:28:10.632 --> 00:28:15.745
talk about ah another issue witches transparency if we are using machine translation in

340
00:28:15.745 --> 00:28:20.493
situation where were saying we have no choice because the

341
00:28:20.493 --> 00:28:25.606
alternative is nothing it's important think still to social eyes the idea of

342
00:28:25.606 --> 00:28:30.719
her people that it is machine translation because that allows them to assess

343
00:28:30.719 --> 00:28:35.832
their own kind of risk tolerance for trusting the information that thereby that thereby

344
00:28:35.832 --> 00:28:40.580
being presented with so transparency is another kind of issue that goes along.

345
00:28:40.580 --> 00:28:44.450
if the ah the risk assessment idea.

346
00:28:44.450 --> 00:28:48.990
i'll just mention at the at the end of the slide ah

347
00:28:48.990 --> 00:28:53.990
little sort of warning that's when you're using free and online

348
00:28:53.990 --> 00:28:58.060
machine translation tools such as google translate but <unk>

349
00:28:58.060 --> 00:29:03.520
the only one when you're using free on mine machine translation usually the terms

350
00:29:03.520 --> 00:29:08.260
and conditions of using that tool include that they may

351
00:29:08.260 --> 00:29:13.620
keep your data and do other things with your data like maybe they'll use it for more training

352
00:29:13.620 --> 00:29:18.620
or enow for some other up purpose that they have aught so it's

353
00:29:18.620 --> 00:29:23.230
good idea not to put confidential information into an on lined machine

354
00:29:23.230 --> 00:29:28.390
transition to make you might think that when you close the window it just disappears

355
00:29:28.390 --> 00:29:33.490
but it doesn't disappear bugle how's your information now and they can use it

356
00:29:33.490 --> 00:29:38.400
according to the terms and conditions that they've laid out which none of us ever read so

357
00:29:38.400 --> 00:29:43.780
we just have to be little bit careful don't put your banking information into an on languishing

358
00:29:43.780 --> 00:29:48.380
transition tool don't put merely personal information that you don't want to be out in the world

359
00:29:48.380 --> 00:29:55.400
because it could it could be out in the world once you've put it in free on languishing translation tall

360
00:29:55.400 --> 00:29:58.090
caney.

361
00:29:58.090 --> 00:30:03.270
wanted to raised the very specific question because we're here

362
00:30:03.270 --> 00:30:07.910
in university context and work at the university context back in canada question

363
00:30:07.910 --> 00:30:12.860
that's getting ocelot is cannot use machine translation for my

364
00:30:12.860 --> 00:30:17.560
course work hut where does not fall in the risk assessment of things

365
00:30:17.560 --> 00:30:22.430
and it's little bit of gray area think on

366
00:30:22.430 --> 00:30:26.730
students are using sheen translation many language teachers wished

367
00:30:26.730 --> 00:30:31.390
that they wouldn't use machine translation but where does the word we kind of lie

368
00:30:31.390 --> 00:30:35.920
on this spectrum and one of the things that i've started to do is sort

369
00:30:35.920 --> 00:30:40.260
of disable what is the learning objective of the course

370
00:30:40.260 --> 00:30:43.130
on all courses have the same learning objectives.

371
00:30:43.130 --> 00:30:48.880
so maybe you are an international student you've calm of from another country to

372
00:30:48.880 --> 00:30:53.920
finland to study and you are here studying science for example

373
00:30:53.920 --> 00:30:58.790
and so the object of of your science course is to learn science

374
00:30:58.790 --> 00:31:03.560
that's the point of your science course try to learn the subject matter of science

375
00:31:03.560 --> 00:31:08.200
and so if you are using machine translation because you have to

376
00:31:08.200 --> 00:31:13.190
hand your work in innes the inn and finish or in english and that's

377
00:31:13.190 --> 00:31:18.470
not the language that you normally speak you might say while i'm using it because

378
00:31:18.470 --> 00:31:23.490
want to annul era want to submit my work to my professor

379
00:31:23.490 --> 00:31:28.250
on i've i've learned the concepts of science and dismayed some help expressing

380
00:31:28.250 --> 00:31:33.210
them so to my mind this is very different objective then being in low

381
00:31:33.210 --> 00:31:40.020
language course where in the language course the objective is to learn the language so maybe french

382
00:31:40.020 --> 00:31:45.770
um and so think that it's you know question of

383
00:31:45.770 --> 00:31:50.150
what is the purpose of the coarse and hurrying on us certainly how does the

384
00:31:50.150 --> 00:31:55.380
teacher feel about it om but in either case think what were doing is drawing

385
00:31:55.380 --> 00:32:00.260
to use machine translation as sort of writing ain't rather than necessarily us

386
00:32:00.260 --> 00:32:05.260
something that will do the whole translation task for us so if you're the science student

387
00:32:05.260 --> 00:32:10.110
and you are just trying to hand in your work in english and demonstrate that you have an understanding of the

388
00:32:10.110 --> 00:32:15.280
concepts one of the things that you'd need to work on is writing your source text

389
00:32:15.280 --> 00:32:20.150
very very well so you're writing the source text in your own language but

390
00:32:20.150 --> 00:32:25.650
you know this isn't an obvious thing to translators but less obvious to people who are translators

391
00:32:25.650 --> 00:32:30.130
if your source text is badly written the translation is going to be disaster current

392
00:32:30.130 --> 00:32:35.060
so you haf to really work on creating good source text in

393
00:32:35.060 --> 00:32:39.960
your own language in order for it to be well translated into another language

394
00:32:39.960 --> 00:32:44.700
so this is actually helping you think to kind of learned the concepts

395
00:32:44.700 --> 00:32:49.200
and learned near the learning objects of the courses to learn science so

396
00:32:49.200 --> 00:32:53.940
you have to learn to write science really well in your own language in order to be able to

397
00:32:53.940 --> 00:32:58.970
benefit from machine translation and hopefully it miching translation is

398
00:32:58.970 --> 00:33:03.530
you know something atoll that you will become last reliant on overtime but don't think

399
00:33:03.530 --> 00:33:08.130
that goes against the objectives of the course in language learning

400
00:33:08.130 --> 00:33:13.510
um what usually advice students is to try it yourself first

401
00:33:13.510 --> 00:33:17.870
and use the machine translation more as verification tool sewed see

402
00:33:17.870 --> 00:33:22.080
how you would do it and then you can compare your results to machine translation

403
00:33:22.080 --> 00:33:26.630
twelve and treated as writing aid rather than as you off

404
00:33:26.630 --> 00:33:31.130
loading the entire task on to machine translation system

405
00:33:31.130 --> 00:33:36.520
either way though think it's important to be respect full of your teachers

406
00:33:36.520 --> 00:33:41.150
instructions if they absolutely do not want you to use it then you should respect

407
00:33:41.150 --> 00:33:46.220
what they've asked of you what try to do now is to say to my students i'm going

408
00:33:46.220 --> 00:33:51.140
to let you use machine translation for many different activities because it's

409
00:33:51.140 --> 00:33:56.290
out their professionals or using it seemed silly not to sing silly not to learn how to use

410
00:33:56.290 --> 00:34:01.480
it in work while with it on occasion there may be something want you to learn and would appreciate

411
00:34:01.480 --> 00:34:06.370
that you not use machine translation right away and so think the students are more receptive

412
00:34:06.370 --> 00:34:11.180
to that because it's not outright forbidden and they know that if i'm asking him not to use it

413
00:34:11.180 --> 00:34:16.190
for particular task that have well thought out reason so rather than blanket kind

414
00:34:16.190 --> 00:34:21.150
of policy yes or no it's sort of it depends on what were doing sniffed

415
00:34:21.150 --> 00:34:26.080
i'm but do asked students always to be transparent so if you used machine translation

416
00:34:26.080 --> 00:34:31.100
for your assignment put it tommy tommy which to use tommy which parts of the text

417
00:34:31.100 --> 00:34:35.800
used it for and show me clearly how you were able to improve it

418
00:34:35.800 --> 00:34:40.490
because chances are they can be improved even if you are using it to get the

419
00:34:40.490 --> 00:34:45.310
gist of something omnino chances are that that you as the

420
00:34:45.310 --> 00:34:50.490
translator trainee or language student are going to be able to recommend someways improving

421
00:34:50.490 --> 00:34:56.250
it because in we just do have operate on different novel

422
00:34:56.250 --> 00:34:59.770
and of course if you're using it tom

423
00:34:59.770 --> 00:35:05.510
in an essay to take the ideas that you found in another paper

424
00:35:05.510 --> 00:35:09.960
on used only to sight those ideas even if your changing the language in which there being

425
00:35:09.960 --> 00:35:15.800
expressed translating something doesn't absolves you of the need to attribute properly

426
00:35:15.800 --> 00:35:20.040
ah through reverencing him citation so those are some questions that have come

427
00:35:20.040 --> 00:35:25.450
up in in art my kind of language teaching environment hands so obvious

428
00:35:25.450 --> 00:35:30.750
that sort of the vice that i'm giving that will be interesting to have discussion

429
00:35:30.750 --> 00:35:35.320
and see what others are doing or what other ideas people have alm mostly

430
00:35:35.320 --> 00:35:40.230
though would say that you know forbidding it doesn't work calm in fact it soared makes it

431
00:35:40.230 --> 00:35:45.710
more enticing and so don't think it makes lot of sense to forbid machine translation

432
00:35:45.710 --> 00:35:50.340
at this point its here an it's not going away the students are using it whether we think

433
00:35:50.340 --> 00:35:55.970
they should are not so see my responsibility as helping him use it in smarter way rather

434
00:35:55.970 --> 00:36:01.470
than trying to arm cunnel locked on the use of machine transaction

435
00:36:01.470 --> 00:36:06.170
so that's the end of kind if the crash corson machine

436
00:36:06.170 --> 00:36:10.890
translation literacy part soul ah wonder if it make sense to take

437
00:36:10.890 --> 00:36:14.720
if there's any questions on that part to take them now and then i'll talk little

438
00:36:14.720 --> 00:36:19.130
bit more about my teaching experience in the second part

439
00:36:19.130 --> 00:36:20.930
all that.

440
00:36:20.930 --> 00:36:26.011
ah do we have any idea where good gulf and other ah

441
00:36:26.011 --> 00:36:31.093
machine translation companies get their parallel corpora from yes would say

442
00:36:31.093 --> 00:36:36.174
that they get it from wide variety of sources hum probably

443
00:36:36.174 --> 00:36:41.255
they do lot of what we might describe scrapings from

444
00:36:41.255 --> 00:36:46.337
the internet of anything that is sort of in the free domain

445
00:36:46.337 --> 00:36:51.418
so many of the european union and the canadian government web sights

446
00:36:51.418 --> 00:36:56.499
are multi lingual for example homme you know lots of into international

447
00:36:56.499 --> 00:37:01.581
or multi natural companies would be multi lingual have have their web

448
00:37:01.581 --> 00:37:06.662
sights in different languages so suspect that there's lot of

449
00:37:06.662 --> 00:37:11.320
us scraping of data from ha hum front from existing free.

450
00:37:11.320 --> 00:37:17.190
jane ascites ah but also everything that gets entered into

451
00:37:17.190 --> 00:37:21.930
the google machine translation system and translated is probably captured

452
00:37:21.930 --> 00:37:27.500
and stored and used adda further level of training so

453
00:37:27.500 --> 00:37:32.110
the more we use the systems were actually generating data for them

454
00:37:32.110 --> 00:37:38.000
come and you may have noticed that some of the tools actually give you the option of editing

455
00:37:38.000 --> 00:37:43.230
little bit improving the translation so they love to see that as well as your kind of doing some

456
00:37:43.230 --> 00:37:48.730
work for them in making the the translations better and so got its it gets

457
00:37:48.730 --> 00:37:54.200
little bit of cycle that feeds itself to some extent mlle

458
00:37:54.200 --> 00:37:59.790
for or ah

459
00:37:59.790 --> 00:38:04.790
your views are very marked resource roars out of him who

460
00:38:04.790 --> 00:38:09.280
are flared while the cell exactly yes that's what mean it's

461
00:38:09.280 --> 00:38:14.630
sort of feeding itself little bet on some of the most active in current

462
00:38:14.630 --> 00:38:19.400
area of machine transition researches on what they call so pathetic data

463
00:38:19.400 --> 00:38:24.350
so it's dated at little bit artificial sort could mean that it was sad translated by machine

464
00:38:24.350 --> 00:38:29.840
itself or generated somehow through another type of ai tool so

465
00:38:29.840 --> 00:38:34.690
absolutely yes what were seeing as that machine translation is being trained with

466
00:38:34.690 --> 00:38:39.380
machine translation data gap to some extent boom

467
00:38:39.380 --> 00:38:44.910
with yours or joke he aright so the next part of the taught the

468
00:38:44.910 --> 00:38:50.230
but i'd like to share with you as little bit about my experience of trying

469
00:38:50.230 --> 00:38:55.780
to develop and and teach machine translation literacy to people who are not

470
00:38:55.780 --> 00:39:00.850
translators or even necessarily ah language students or were

471
00:39:00.850 --> 00:39:06.130
having strong background informal language learning

472
00:39:06.130 --> 00:39:07.980
hang me.

473
00:39:07.980 --> 00:39:13.100
ah so when talk about machine translation literacy what

474
00:39:13.100 --> 00:39:18.190
i'm really trying to describe as sort of type of the digital plus

475
00:39:18.190 --> 00:39:23.030
information literacy that focuses on developing critical thinking

476
00:39:23.030 --> 00:39:28.150
skills rather than technical competence ain't so it's not really about

477
00:39:28.150 --> 00:39:33.340
how to use machine translation in the sense of which button to push

478
00:39:33.340 --> 00:39:38.220
because really the technologies super simple right it's it's sort of like copy

479
00:39:38.220 --> 00:39:43.280
paced click and that's all there is to using it technically i'm

480
00:39:43.280 --> 00:39:47.940
like some other take to suffer which are very sophisticated and really require lot of effort

481
00:39:47.940 --> 00:39:53.300
to learn how to use them in gigantic user manual that you have to read machine translation

482
00:39:53.300 --> 00:39:57.860
as user is very simple saw for two years

483
00:39:57.860 --> 00:40:02.890
i'm so it's not about how enow to learn which button to push but it's really

484
00:40:02.890 --> 00:40:08.640
about making decisions such as nor should be using machine translation in this context

485
00:40:08.640 --> 00:40:13.310
what are the consequences of using it in this context alm what

486
00:40:13.310 --> 00:40:19.010
might use instead of rushing translation or no should use machine translation

487
00:40:19.010 --> 00:40:23.130
this time but not next time so it's really about this sort of cognitive

488
00:40:23.130 --> 00:40:28.110
i'm thinking that goes around machine translation use rather than

489
00:40:28.110 --> 00:40:33.210
the technical um side of it we might also think about how

490
00:40:33.210 --> 00:40:38.250
we as people could interact with the machine transition system to improve

491
00:40:38.250 --> 00:40:43.050
the output and as said sometimes that interaction might be before we give tax to

492
00:40:43.050 --> 00:40:48.000
the machine transition sayst sometimes it might be taking the output of the machine transition system

493
00:40:48.000 --> 00:40:53.110
in trying to make it better before we use it for certain purpose so what is our role

494
00:40:53.110 --> 00:40:58.030
in the process and how can we worked together with the machine

495
00:40:58.030 --> 00:41:03.190
and so by asking this type of question we sort of have the potential to become

496
00:41:03.190 --> 00:41:08.230
more informed and more critical users instead of just doing this autopilot

497
00:41:08.230 --> 00:41:13.290
where we copy paced click done that next i'm here

498
00:41:13.290 --> 00:41:18.420
but do have to say that in when first are to think me but wishing transition literacy asserted

499
00:41:18.420 --> 00:41:23.980
thought of it as as one concept but now realized the vastly

500
00:41:23.980 --> 00:41:28.750
different types of users who were out there and the different purposes

501
00:41:28.750 --> 00:41:33.270
that they might have for using machine translation now see machine translation literacy

502
00:41:33.270 --> 00:41:38.470
as much more of kind of custom eyes abul concept language students might need

503
00:41:38.470 --> 00:41:43.320
different information about wishing translation than on in people who are using

504
00:41:43.320 --> 00:41:48.270
it for leisure purposes versus people who are trying to use it in their work placed so

505
00:41:48.270 --> 00:41:53.110
old machine translation literacy just like machine translation use might look different

506
00:41:53.110 --> 00:41:57.700
so i'm talking about my experience to day which is mostly working with

507
00:41:57.700 --> 00:42:02.590
students at universities om but students who are not translation

508
00:42:02.590 --> 00:42:07.040
students of that's my context but you might want to adjust machine translation literacy

509
00:42:07.040 --> 00:42:11.540
little bit depending on the type of people that you're working with

510
00:42:11.540 --> 00:42:14.150
tom okay.

511
00:42:14.150 --> 00:42:19.380
so firstly what's driving the need for machine physician

512
00:42:19.380 --> 00:42:24.650
literacy wire we even talking about this here to day dam and of course one reason

513
00:42:24.650 --> 00:42:29.630
i've already alluded to briefly is that machine translation is now

514
00:42:29.630 --> 00:42:34.310
out there right for many years starting from the time during

515
00:42:34.310 --> 00:42:39.310
the post second world war up to around the year two thousand or even two thousand

516
00:42:39.310 --> 00:42:44.440
five miching translation wasn't freely available it was an easy to get your

517
00:42:44.440 --> 00:42:49.360
hands on it was only researches and developers and may be professional translators that

518
00:42:49.360 --> 00:42:54.470
had access to the psychology but in the last fifteen years that has been turned upside

519
00:42:54.470 --> 00:42:59.600
down completely by tools like google translate now anyone can get their hands

520
00:42:59.600 --> 00:43:04.430
on machine translation all you need is internet connection and you've got machine translation of

521
00:43:04.430 --> 00:43:06.020
hustle to you.

522
00:43:06.020 --> 00:43:09.130
so that's one big changed.

523
00:43:09.130 --> 00:43:15.040
the second choice the everything is hopefully numbered one here or about the

524
00:43:15.040 --> 00:43:20.060
this second um item is that again as mention machine translations

525
00:43:20.060 --> 00:43:25.180
very easy to use and in some ways it almost too easy because when something is

526
00:43:25.180 --> 00:43:30.750
super easy to do we don't think about it which is do it and like on autopilot

527
00:43:30.750 --> 00:43:35.810
so the fact that it's so simple to use up puts us in risky

528
00:43:35.810 --> 00:43:41.140
situation where we don't necessarily take the time to think about whether we should be using it

529
00:43:41.140 --> 00:43:46.200
ah we've also talked about the data driven approached machine translation and one of

530
00:43:46.200 --> 00:43:51.280
the big differences between the output of dated driven approaches

531
00:43:51.280 --> 00:43:56.210
to machine translation and those old rule based approaches to machine translation is

532
00:43:56.210 --> 00:44:01.120
that the output of machine translation using dated

533
00:44:01.120 --> 00:44:06.280
driven approach sounds good the older approaches produced really

534
00:44:06.280 --> 00:44:11.790
clung key sentences and it was actually quite easy to tell

535
00:44:11.790 --> 00:44:16.410
that it was machine translation because it just sounded weird nay it sounded clung

536
00:44:16.410 --> 00:44:21.630
key and we might even collot sort of translation needs you know when you're reading about translation

537
00:44:21.630 --> 00:44:26.300
it just sounds off the words might be in the language

538
00:44:26.300 --> 00:44:31.090
but there and they don't go together well unjust sounds weird

539
00:44:31.090 --> 00:44:36.660
with driven machine translation it sounds good even if it's not write

540
00:44:36.660 --> 00:44:41.690
it sounds good so at last us into this fall sense of security it sounds

541
00:44:41.690 --> 00:44:47.270
reasonable so it must be right and sometimes it's not correct so it's harder

542
00:44:47.270 --> 00:44:52.430
to spot the errors and the errors are different from what we were used to seeing

543
00:44:52.430 --> 00:44:57.360
with the older approach astonishing translation so that the challenge

544
00:44:57.360 --> 00:45:02.660
and something else says that the popular media really seems to like talking about machine

545
00:45:02.660 --> 00:45:08.040
translation now it's thoughtfully up fairly regular topic in the popular media

546
00:45:08.040 --> 00:45:13.410
but the problem is that the popular media tend to present sort of binary

547
00:45:13.410 --> 00:45:17.980
um description of machine translation either

548
00:45:17.980 --> 00:45:22.770
they are picking on the mistakes and sort of saying that his

549
00:45:22.770 --> 00:45:27.710
terrible it's horrible tool or there really high ping and up at the

550
00:45:27.710 --> 00:45:32.740
other and and sing like waugh pretty soon we were even need translators like these are amazing their

551
00:45:32.740 --> 00:45:38.120
practically magic and those two extremes are

552
00:45:38.120 --> 00:45:42.770
not the reality for most machine collision which is really full summer in the

553
00:45:42.770 --> 00:45:47.770
middle so we have to be little bit careful about what we read in popular media because

554
00:45:47.770 --> 00:45:53.080
not always is of the case that the journalist understands the technology so they're not presenting

555
00:45:53.080 --> 00:45:58.130
it necessarily in the best uppermost new wonst light

556
00:45:58.130 --> 00:45:59.900
bouquet.

557
00:45:59.900 --> 00:46:05.190
so if all of those things were happening how we don't have ha background

558
00:46:05.190 --> 00:46:10.710
in translation the tools are really easy to use the medius telling us that that their amazing

559
00:46:10.710 --> 00:46:15.040
up or their terrible amina what where were does this leaf

560
00:46:15.040 --> 00:46:19.980
people wider we think that people should instinctively know how to use

561
00:46:19.980 --> 00:46:25.150
machine translation in smart way ah and the answer is that that they don't

562
00:46:25.150 --> 00:46:30.330
with some one who has no background in translation know how to use transition thrill

563
00:46:30.330 --> 00:46:35.160
in smart way it's like have no background in medicine don't let me loose

564
00:46:35.160 --> 00:46:40.070
in an operating theatre and let me start like no operating on people that would

565
00:46:40.070 --> 00:46:45.180
be terrible idea so why do we think that people who have no background in translation should

566
00:46:45.180 --> 00:46:50.020
just know how to use machine translation in smart way they don't and so

567
00:46:50.020 --> 00:46:54.220
feel that we in the language professions of the language industry have little bit

568
00:46:54.220 --> 00:46:58.680
of social responsibility to help them ah i'm develop good

569
00:46:58.680 --> 00:47:02.520
critical thinking skills around machine translation

570
00:47:02.520 --> 00:47:08.420
so i've started programme of machine transition literacy at my own university

571
00:47:08.420 --> 00:47:12.570
and in hope it basically looks like the first half of the presentation

572
00:47:12.570 --> 00:47:17.630
to day where we talk little bit about how machine transition works what are the

573
00:47:17.630 --> 00:47:22.750
the in kind of implications of the data driven approach to machine translation on the fact that

574
00:47:22.750 --> 00:47:27.730
are our multiple machine physician systems out there that we can try we talk

575
00:47:27.730 --> 00:47:32.710
little bit about this risk assessment idea which tasks are high stakes which

576
00:47:32.710 --> 00:47:37.740
are low stakes and how can we developed good judgment about when to use machine

577
00:47:37.740 --> 00:47:43.280
translation we talk about transparency that important to

578
00:47:43.280 --> 00:47:47.640
be honest if if you're using machine has ation tool you don't need to be ashamed

579
00:47:47.640 --> 00:47:52.650
of it's not shameful to use machine translation but eugenie to be transparent about it because that

580
00:47:52.650 --> 00:47:58.040
allows the recipient of the translated information to determine their own

581
00:47:58.040 --> 00:48:03.100
kind of comfort level in trusting that information came and we tottleben

582
00:48:03.100 --> 00:48:08.060
about interacting with machine translation the this idea of garbage in equals

583
00:48:08.060 --> 00:48:13.070
garbage out right that if you want good translation you need to provide good source

584
00:48:13.070 --> 00:48:18.100
text om and that at the other end to know the though translation that's

585
00:48:18.100 --> 00:48:23.030
produced by the machine might not be perfect and so depending on what

586
00:48:23.030 --> 00:48:28.470
you're planning to do with that information you might need to improve the text he might mean to do some editing

587
00:48:28.470 --> 00:48:33.160
were arranged for someone else to do the editing if you can to yourself but it may

588
00:48:33.160 --> 00:48:38.360
not be enough to does stopped with the raw machine translation out so that the lykon

589
00:48:38.360 --> 00:48:43.180
of type of thing that talked about in machine transition literacy as

590
00:48:43.180 --> 00:48:47.380
said it's custom icicle concept you might find that the groups that you're working with

591
00:48:47.380 --> 00:48:51.550
would benefit from other information in and thoughts of terrific you

592
00:48:51.550 --> 00:48:55.720
might be well placed to determine what that should be

593
00:48:55.720 --> 00:49:01.320
ha hum so what i've actually started awfu

594
00:49:01.320 --> 00:49:06.450
is to pilot of course on translation for non translators

595
00:49:06.450 --> 00:49:10.820
and machine transition literacy is just one part of that so

596
00:49:10.820 --> 00:49:15.810
it's not shah <unk> twelve we course on machine translation literacy it's

597
00:49:15.810 --> 00:49:20.810
twelve what course on translation and machine translation literacy as one model within

598
00:49:20.810 --> 00:49:26.110
that so is the only ma joined in talk about today but didn't want to giving the impression that the whole twelve

599
00:49:26.110 --> 00:49:31.040
weeks is about wishing translation hum but it's course that is

600
00:49:31.040 --> 00:49:36.220
the ah knew and our university anyway and its of course that is basically

601
00:49:36.220 --> 00:49:40.980
aimed at every one except people in the professional translation programme

602
00:49:40.980 --> 00:49:45.880
so it's first your course an it's very am open it has no pre requisite

603
00:49:45.880 --> 00:49:51.100
it's and students from any other faculty can register to take course so

604
00:49:51.100 --> 00:49:56.380
we've got students who were coming from the education faculty from engineering hell sciences

605
00:49:56.380 --> 00:50:00.980
management science social science all in all so far this

606
00:50:00.980 --> 00:50:05.930
is of the force iteration of the course so ice first taught it of him in

607
00:50:05.930 --> 00:50:10.910
the false master last year sewing in the fourth iteration of the coarse now

608
00:50:10.910 --> 00:50:16.390
and seen students from forty seven different programs of the university come through so it's

609
00:50:16.390 --> 00:50:20.940
really diverse the group whose there and also asked them

610
00:50:20.940 --> 00:50:26.030
to sort of self declare how om we have in canada of course to official languages

611
00:50:26.030 --> 00:50:31.140
french and english but we get many international students we have many students who

612
00:50:31.140 --> 00:50:35.880
have heritage language maybe they or their parents had emigrated from another country

613
00:50:35.880 --> 00:50:41.380
the soul we see diverse range of languages i've asked just

614
00:50:41.380 --> 00:50:45.980
enough students to self declare what their arm native or dominant languages

615
00:50:45.980 --> 00:50:51.540
and their friend twenty eight different languages in the course of heart

616
00:50:51.540 --> 00:50:56.080
and soul we have seen approximately two hundred

617
00:50:56.080 --> 00:51:01.130
students have gone through the course so far and asked them to do little survey on

618
00:51:01.130 --> 00:51:06.730
the machine translation literacy part of the coarse and i'faith agata hundred and seventy responses

619
00:51:06.730 --> 00:51:11.420
soul around eighty five percent response rate soit swinish share with you little bit

620
00:51:11.420 --> 00:51:16.160
ah what the results of of thought us survey were

621
00:51:16.160 --> 00:51:21.170
and what is the first questions but ask the students is how often do

622
00:51:21.170 --> 00:51:26.650
use machine translation is it like something that's regular for you or hardly ever

623
00:51:26.650 --> 00:51:31.050
and you can see that on you know up all over eighty per

624
00:51:31.050 --> 00:51:36.100
cent of the students are using it at least once month and many of them are using it much

625
00:51:36.100 --> 00:51:40.890
more often than that sometimes even every day so they are definitely using

626
00:51:40.890 --> 00:51:45.870
the technology it's not an alien concept to them it's not os something that

627
00:51:45.870 --> 00:51:50.620
they haena might use once year there using it quite regularly

628
00:51:50.620 --> 00:51:55.040
and there not translators made the students are not translation students

629
00:51:55.040 --> 00:51:59.440
of their of people who are doing in studying other things

630
00:51:59.440 --> 00:52:05.020
and asked them fino in what area of your life

631
00:52:05.020 --> 00:52:09.520
do use machine translation so of course i'm talking to students it's not terribly

632
00:52:09.520 --> 00:52:14.480
surprising that many of them are using it for their coarse work because your that's kind of their what

633
00:52:14.480 --> 00:52:19.400
they'd spent most of their time doing or studying and and so there it might be using machine

634
00:52:19.400 --> 00:52:24.790
translation for their courses but they also use it some of them might have part time job outside

635
00:52:24.790 --> 00:52:29.320
of their studies and the using it for leisure activities as well

636
00:52:29.320 --> 00:52:34.220
as butts of their life there using wishing consolation

637
00:52:34.220 --> 00:52:39.500
was also interested to know whether

638
00:52:39.500 --> 00:52:45.190
they use machine translation mostly to kind of absorb information

639
00:52:45.190 --> 00:52:49.450
so there starting with text in language that they don't know and translating

640
00:52:49.450 --> 00:52:54.320
it into their own language or whether they're trying to use machine translation to help them

641
00:52:54.320 --> 00:52:59.440
produce information to share with other people and it seems that

642
00:52:59.440 --> 00:53:04.320
most students use it in both directions so they are using machine translation both

643
00:53:04.320 --> 00:53:09.350
to kind of assimilate ah information in it from another language but also to

644
00:53:09.350 --> 00:53:14.490
disseminate information so these are actually two very different tasks and

645
00:53:14.490 --> 00:53:19.360
machine translation may be better or worse first for some era

646
00:53:19.360 --> 00:53:24.250
of those purposes so it's interesting to know that students are using it mainly in both

647
00:53:24.250 --> 00:53:26.100
directions.

648
00:53:26.100 --> 00:53:31.140
also wanted to get field for before

649
00:53:31.140 --> 00:53:35.740
they learn anything about in machine physician literacy how happy are they with machine

650
00:53:35.740 --> 00:53:40.330
translation to they think that its working well to meet the needs that they have

651
00:53:40.330 --> 00:53:45.200
and most students are reasonably satisfied him very few are completely

652
00:53:45.200 --> 00:53:49.570
satisfied very few are completely dissatisfied most are coming down

653
00:53:49.570 --> 00:53:54.260
somewhere in the middle either very or moderately satisfied

654
00:53:54.260 --> 00:53:59.420
so that gives me hope thought

655
00:53:59.420 --> 00:54:04.220
tom you know there if they could learn something for machine transition illiteracy and perhaps

656
00:54:04.220 --> 00:54:09.130
opfer subsection level little bit by learning how to use it sometimes it's not that the

657
00:54:09.130 --> 00:54:13.850
the tuneless bowed but that we've used it for the wrong thing or that we having shown

658
00:54:13.850 --> 00:54:18.760
good judgment around it so i'm hopeful that this hour result means that

659
00:54:18.760 --> 00:54:23.700
they're not so disappointed in it that they're goin to give up on it but there's room for improvement

660
00:54:23.700 --> 00:54:28.160
to be able to use it the it better more effectively

661
00:54:28.160 --> 00:54:33.320
ah big question how

662
00:54:33.320 --> 00:54:38.340
asked them if there was no free on languishing translation system what would

663
00:54:38.340 --> 00:54:44.060
you do instead and the really interesting thing for me here is that nobody

664
00:54:44.060 --> 00:54:48.530
would pay professional now we are talking about students students

665
00:54:48.530 --> 00:54:53.970
don't have lot of money on but the types of needs that students hof

666
00:54:53.970 --> 00:54:58.520
they would not pay professional and this is really important because

667
00:54:58.520 --> 00:55:03.840
in the transition community high think there's often quite lot of concern that machine

668
00:55:03.840 --> 00:55:09.070
translation is taking away business from professional translators her this group

669
00:55:09.070 --> 00:55:13.810
would not be paying professional translators anyway so there's really no need for

670
00:55:13.810 --> 00:55:18.700
us or to false competition between professional translators and

671
00:55:18.700 --> 00:55:24.120
inch of machine translation systems at least not in the context of students

672
00:55:24.120 --> 00:55:29.700
most of them said all i'd ask friends to help or colleague om couple would translate

673
00:55:29.700 --> 00:55:34.900
effects themselves ah i'm quite if you would just do nothing so it's

674
00:55:34.900 --> 00:55:40.230
it's really not ah this use case anyway is not threat to professional

675
00:55:40.230 --> 00:55:46.300
translation that this is type of translation where the needs were simply going on met before

676
00:55:46.300 --> 00:55:50.950
on its it's not business that's being taken away from translators

677
00:55:50.950 --> 00:55:55.980
when asked them how out of all the different things

678
00:55:55.980 --> 00:56:01.230
that we learned in the machine transition literacy ma jewel which one surprised

679
00:56:01.230 --> 00:56:06.310
you the most or what was the most novel thing that he learns that you hadn't expected about wishing

680
00:56:06.310 --> 00:56:11.410
her violation and the most common one was plum they were shocked to find

681
00:56:11.410 --> 00:56:16.080
out said that that google or the other transition companies could keep their data

682
00:56:16.080 --> 00:56:21.260
and an lot of soon saw unfit to be little more careful now about what enter into

683
00:56:21.260 --> 00:56:26.200
machine hesitation system it never occurred to me i'll or lot of students said it didn't occur

684
00:56:26.200 --> 00:56:31.240
to me but now that you pointed out or think yacht of course there keeping my data what was what

685
00:56:31.240 --> 00:56:36.760
was to be thinking but come yet that was the most common one but quite variety

686
00:56:36.760 --> 00:56:41.200
as you see so that again made me think that machine translation literacy

687
00:56:41.200 --> 00:56:47.100
is worth it because they are learning different things some people knew acts some people knew why but

688
00:56:47.100 --> 00:56:52.950
nobody knew everything and almost every one learns something new in machine translation literacy mont

689
00:56:52.950 --> 00:56:55.040
to me.

690
00:56:55.040 --> 00:56:59.850
and ha finally

691
00:56:59.850 --> 00:57:04.730
asked them to questions at the and said you are not

692
00:57:04.730 --> 00:57:10.110
translator you're not planning to become translator but you do use machine translation

693
00:57:10.110 --> 00:57:14.530
do you think that this montell unashamed translation literacy

694
00:57:14.530 --> 00:57:19.210
or the concept of machine transition literacy in general is important for people who

695
00:57:19.210 --> 00:57:24.090
are not planning to become language professionals and most

696
00:57:24.090 --> 00:57:28.870
of the students said yes it was very important they really felt that they had learnt something

697
00:57:28.870 --> 00:57:33.520
and that it would change their approach to using machine translation in the future

698
00:57:33.520 --> 00:57:38.570
hardly anyone set it wasn't worth while whom no not everyone thought it was super important

699
00:57:38.570 --> 00:57:43.820
but most people thought it was no it was worth learning they had de felt it was time whilst

700
00:57:43.820 --> 00:57:48.870
that and the very last question said was do you think then

701
00:57:48.870 --> 00:57:53.680
based on your experience that this because it's pilot project at the moment do

702
00:57:53.680 --> 00:57:59.070
you think that we should make this of permanent part of the universities offering

703
00:57:59.070 --> 00:58:03.660
and that we should be teaching machine translation literacy to all students even

704
00:58:03.660 --> 00:58:09.150
the ones who aren't planning look of career in the language professions and again was really

705
00:58:09.150 --> 00:58:13.760
arm pleasantly surprised by the kind of positive um

706
00:58:13.760 --> 00:58:18.960
encouragement for doing us for saying yes it does seem to be worth while people who art

707
00:58:18.960 --> 00:58:23.700
in the length refreshments are still users of this technology and are perhaps more in

708
00:58:23.700 --> 00:58:27.980
need of it because they won't get it this type of education anywhere else.

709
00:58:27.980 --> 00:58:32.380
so that was the last question in the survey which

710
00:58:32.380 --> 00:58:36.130
so just to wrap things up has sort of take away

711
00:58:36.130 --> 00:58:41.410
ah if could say anything about machine translation would

712
00:58:41.410 --> 00:58:46.830
say it is not going to it is here it's free it's easy

713
00:58:46.830 --> 00:58:52.020
to use of course people are going to use it they are using it as said

714
00:58:52.020 --> 00:58:57.130
the i'm reason estimate from google themselves was that there

715
00:58:57.130 --> 00:59:02.670
are billion users of this technology so even if we think no

716
00:59:02.670 --> 00:59:08.000
and by we i'm in people in the language professions think that people outside length professions

717
00:59:08.000 --> 00:59:13.260
shouldn't useless technology that they should turn to professionals for high quality service

718
00:59:13.260 --> 00:59:18.130
instead it doesn't matter what we think they are using it

719
00:59:18.130 --> 00:59:23.370
and there using it often enter using it for of variety of purposes

720
00:59:23.370 --> 00:59:28.320
so think that we should be left concerned about this idea

721
00:59:28.320 --> 00:59:33.330
of competing hum with machine translation ah many of the

722
00:59:33.330 --> 00:59:38.480
things that people use it for they would not have hired professional translator for anyway said

723
00:59:38.480 --> 00:59:43.820
not taking away any business and to this pop point

724
00:59:43.820 --> 00:59:48.800
is think also really important things that are obvious to translators are not

725
00:59:48.800 --> 00:59:53.610
obvious to non translators and echoes backed to to my point sounded like

726
00:59:53.610 --> 00:59:58.910
why would we think that people who don't have background of preparation training

727
00:59:58.910 --> 01:00:03.700
in translation would be able to use the stool in the smartest way possible

728
01:00:03.700 --> 01:00:08.680
ah unless we helped them so do think that there is sort of social responsibility on

729
01:00:08.680 --> 01:00:13.870
on the part of the language professions because think we has been guilty of little bit

730
01:00:13.870 --> 01:00:18.200
of slamming the technology saying that it's really

731
01:00:18.200 --> 01:00:23.610
not goulden you should be known your you really mean to be careful

732
01:00:23.610 --> 01:00:27.700
you should be hiring professional translator instead and think that we need

733
01:00:27.700 --> 01:00:32.460
to change or tone little bit and recognized that caught in week

734
01:00:32.460 --> 01:00:37.220
can play roland helping people and that it's not harming us helping others as

735
01:00:37.220 --> 01:00:41.930
not harming us and so am think that that made me too

736
01:00:41.930 --> 01:00:47.210
think about that little more and i'm actually interested in where and when we should be doing

737
01:00:47.210 --> 01:00:52.200
machine translation literacy interventions on and i've been working

738
01:00:52.200 --> 01:00:57.230
lot with first or university students but it seems to me that tom

739
01:00:57.230 --> 01:01:02.580
it might be worth introducing this to people who are even younger because their arriving at university

740
01:01:02.580 --> 01:01:07.330
already having used machine translation so maybe we should be thinking

741
01:01:07.330 --> 01:01:12.560
about integrating it into high schools or maybe even middle school don't know

742
01:01:12.560 --> 01:01:17.190
am to do that we need to reach an important group and the where i'm

743
01:01:17.190 --> 01:01:22.750
sort of trying to tackle now is the fly dia of odd teaching the teacher's

744
01:01:22.750 --> 01:01:27.460
so we can't necessarily put translator in every class room but if we could get

745
01:01:27.460 --> 01:01:32.130
out involved in may be integrating machine translation literacy education in

746
01:01:32.130 --> 01:01:37.210
to teacher training than the teacher's themselves could carry that forward to the students

747
01:01:37.210 --> 01:01:42.060
and think that am it could be part of sort of media literacy

748
01:01:42.060 --> 01:01:46.230
did illiteracy information literacy kind of package and we know that

749
01:01:46.230 --> 01:01:51.260
they have those at schools are ready with sort of me to inject little bit of machine translation

750
01:01:51.260 --> 01:01:55.710
literacy into that overall package to to reach on much

751
01:01:55.710 --> 01:02:00.120
wider think group of the gentle public

752
01:02:00.120 --> 01:02:04.670
and will end their high will put up just when last were people can

753
01:02:04.670 --> 01:02:09.560
contact me you're interested in knowing more about the machine translation literacy

754
01:02:09.560 --> 01:02:13.870
project we have web sight and were on twitter and you can

755
01:02:13.870 --> 01:02:18.250
of course female me at my home institution

756
01:02:18.250 --> 01:02:20.930
and there's the book of course the <unk>.

757
01:02:20.930 --> 01:02:24.180
faith faith.

758
01:02:24.180 --> 01:02:29.130
oh yes have to say am so i'm

759
01:02:29.130 --> 01:02:34.280
attached to school of translation and interpretation and very much in translator

760
01:02:34.280 --> 01:02:38.780
training so don't princely do language teaching but i'm interested in having more

761
01:02:38.780 --> 01:02:43.850
conversations with people who do language teaching because agree with you think you're parallels

762
01:02:43.850 --> 01:02:48.700
and think that we could the um learned from each other and and perhaps

763
01:02:48.700 --> 01:02:53.290
reach wider audience by working together yah

764
01:02:53.290 --> 01:02:55.130
by.

765
01:02:55.130 --> 01:03:00.880
now really hope that do think it's something that many groups could participate

766
01:03:00.880 --> 01:03:06.060
in hand really loved the language campus sat up that you

767
01:03:06.060 --> 01:03:11.820
have here where you bring people together from the languished apartment but also from education or an also

768
01:03:11.820 --> 01:03:16.910
people with interest and technologies an do think that this is kind of

769
01:03:16.910 --> 01:03:22.130
into disciplinary issue and that we will

770
01:03:22.130 --> 01:03:27.140
in oh do better job of keineth's solving it or tackling it by working

771
01:03:27.140 --> 01:03:32.240
together out because think translation is one facet of it but

772
01:03:32.240 --> 01:03:36.820
its also about teaching its also about which her nerves yacht

773
01:03:36.820 --> 01:03:41.890
bum new critical development judgment which can be

774
01:03:41.890 --> 01:03:46.450
honed through other avenues then only through translation

775
01:03:46.450 --> 01:03:51.580
<unk> <unk> yet do actually strongly

776
01:03:51.580 --> 01:03:57.310
believin that think there's been sort of false idea of competition between

777
01:03:57.310 --> 01:04:01.750
profession translators and machine translation partly mean partly

778
01:04:01.750 --> 01:04:06.800
it developed way back in that period in the following second were were

779
01:04:06.800 --> 01:04:11.850
part of the problem was that at the time turned now there was nothing

780
01:04:11.850 --> 01:04:16.940
like constitutional linguist expire was nothing like language technology at the time

781
01:04:16.940 --> 01:04:21.690
all of the workin machine translation was being done only by those who really computer

782
01:04:21.690 --> 01:04:26.860
scientists to her often mathematicians or electrical engineers and

783
01:04:26.860 --> 01:04:31.770
they thought that miching translation would be an easy thing for them to to

784
01:04:31.770 --> 01:04:36.660
salt they use niecks to solving translation so at the time um

785
01:04:36.660 --> 01:04:41.760
the people working on machine translation were perhaps little bit naive about

786
01:04:41.760 --> 01:04:46.980
what was involved because they were translators themselves they didn't knoll and sold

787
01:04:46.980 --> 01:04:52.190
the message ing that went out in the early years was quite destructive think they really

788
01:04:52.190 --> 01:04:57.640
over promised and then under delivered and soul it left bad

789
01:04:57.640 --> 01:05:01.980
taste really in the kind of hummed mouth of translator to have

790
01:05:01.980 --> 01:05:07.300
been told you know of this technologies can put you out of business that was the initial

791
01:05:07.300 --> 01:05:12.020
massah jack came out and of course we in transition professions realize its much more

792
01:05:12.020 --> 01:05:17.150
complex than code breaking its not something that's an easy fix ham

793
01:05:17.150 --> 01:05:22.080
bite it set bad tore right from the beginning which was unfortunate but think we have enough

794
01:05:22.080 --> 01:05:27.050
distance enough history noun enough opportunities for collaboration that we do need

795
01:05:27.050 --> 01:05:32.240
to change that from such negative tone to seraphina we have valuable

796
01:05:32.240 --> 01:05:36.400
asgill to offer we can help you learnin and why not have you learn

797
01:05:36.400 --> 01:05:40.800
its not it's not harming us so now

798
01:05:40.800 --> 01:05:45.590
ardors ya i'm sure don't personally know

799
01:05:45.590 --> 01:05:50.910
how to do it i'm but i'm quite certain that there are tools out there in their called web scrapers

800
01:05:50.910 --> 01:05:55.040
so c'est

801
01:05:55.040 --> 01:06:00.220
scrape scrape or an to um think if you look into literature

802
01:06:00.220 --> 01:06:05.200
on the digital humanity's think you'll find information about when scraping in the digital

803
01:06:05.200 --> 01:06:10.110
humanity's yacht all there is this lot of especially government

804
01:06:10.110 --> 01:06:15.350
things as lot of free open om like it's not copyrighted material

805
01:06:15.350 --> 01:06:20.250
when its government shall wed sights and things like that it stood open domain and sewed so you

806
01:06:20.250 --> 01:06:25.150
don't need permission you don't need license by sort of thing what

807
01:06:25.150 --> 01:06:30.370
if you'd sit in the freed of open domain you can use and lot of mean governments

808
01:06:30.370 --> 01:06:35.590
of countries that our officially bi lingual will have lot of of that information

809
01:06:35.590 --> 01:06:40.220
on that's very particular text type like your only getting government stuff but it's enough to

810
01:06:40.220 --> 01:06:45.200
play with and too kind of start ya bon

811
01:06:45.200 --> 01:06:50.210
so in the alms scholarly community this is actually becoming

812
01:06:50.210 --> 01:06:55.410
of very active area of interest on what where seeing is thatched

813
01:06:55.410 --> 01:07:00.560
there's push back now up for many years english husband the lingua franca of scholarly

814
01:07:00.560 --> 01:07:05.250
publishing and you know it makes sense on the one hand to

815
01:07:05.250 --> 01:07:10.570
use lingua franca because there are many benefits but on the other hand there also many drawbacks and

816
01:07:10.570 --> 01:07:15.220
the drawbacks are particularly for the ninety five per cent of research was who are not

817
01:07:15.220 --> 01:07:20.280
native speakers of english soul omnino it's it's quite unbalanced

818
01:07:20.280 --> 01:07:25.370
i'm just earlier this month ah you nest go issued recommendation on

819
01:07:25.370 --> 01:07:30.230
open science and one of the things that they recommended was that science should

820
01:07:30.230 --> 01:07:35.230
be more multi lingual that foreseeing all of science through single

821
01:07:35.230 --> 01:07:40.470
common language leads in some cases to poor science and can have very

822
01:07:40.470 --> 01:07:45.560
detrimental affects um and so they are advocating for multi

823
01:07:45.560 --> 01:07:50.450
lingual science which in no it in the problem is that multi

824
01:07:50.450 --> 01:07:55.360
lingual and lingua franca they each have strengths and weaknesses the strength of one is the weakness

825
01:07:55.360 --> 01:08:00.330
of the other so there's no perfect solution omelet but since we have

826
01:08:00.330 --> 01:08:05.700
seen already thee down side of using single language common

827
01:08:05.700 --> 01:08:10.710
language for science on think people are interested in exploring whether

828
01:08:10.710 --> 01:08:15.920
the other option our will be better will solve up some of the problems and

829
01:08:15.920 --> 01:08:20.580
think machine translation is poised to become an important part of that

830
01:08:20.580 --> 01:08:25.380
but it can't be it ever seen as the panacea it's not the solution

831
01:08:25.380 --> 01:08:30.440
and it's not going to fix everything but think can play an important roll had but think

832
01:08:30.440 --> 01:08:36.030
this literacy is equally important because they were going to again be seeing

833
01:08:36.030 --> 01:08:40.510
scholars who are scholars of other disciplines not translation but scholars

834
01:08:40.510 --> 01:08:45.870
of everything else who again why with they know how to use machine translation

835
01:08:45.870 --> 01:08:50.490
sensibly ha some of the challenges that wagon face are ob de l'eau

836
01:08:50.490 --> 01:08:55.890
domain below resource domain because the more specialized our work the fewer articles

837
01:08:55.890 --> 01:09:00.820
that have been written about it omne <unk> the purpose of researches

838
01:09:00.820 --> 01:09:05.710
often innovation meaning nothing has been written about it raises brand new thing

839
01:09:05.710 --> 01:09:10.690
omnes so those or some the challenges the organ half too think take on our if

840
01:09:10.690 --> 01:09:15.540
we want to really see machine translation be successful in unscholarly communication howdah

841
01:09:15.540 --> 01:09:20.620
rebuild the corpora howdah we amass enough data in all the languages that we need and

842
01:09:20.620 --> 01:09:26.260
all the subject feels that we need and think it's going to require cooperation because

843
01:09:26.260 --> 01:09:30.930
the other tonto scully communication right now is that most of the writing

844
01:09:30.930 --> 01:09:35.820
is behind pale wall rate almost all of the journals require subscription

845
01:09:35.820 --> 01:09:41.140
access the only thing it's available as the abstract which is in all something

846
01:09:41.140 --> 01:09:45.770
but it's not enough it's hot so we're going to need collaboration from

847
01:09:45.770 --> 01:09:51.470
the publishing industry to make the data available for corpora

848
01:09:51.470 --> 01:09:56.150
at in order to the have something to feed the machine translation system for it to learn from

849
01:09:56.150 --> 01:10:00.830
so it's it's definitely going to be collaborate of effort and in

850
01:10:00.830 --> 01:10:05.700
all open access is taken off open educational resources open pub

851
01:10:05.700 --> 01:10:11.240
rushing out so that might help lot if we could ha have more people publishing

852
01:10:11.240 --> 01:10:16.060
in open access that they'd have becomes available for us system

853
01:10:16.060 --> 01:10:21.070
such as machine translation on and this was just really really interested by this

854
01:10:21.070 --> 01:10:25.940
hour recommendation from unit score justice months that in we need to

855
01:10:25.940 --> 01:10:31.600
they framed it in way of hum inclusion rate to make science

856
01:10:31.600 --> 01:10:36.650
more inclusive we need to ah have more multi lingual

857
01:10:36.650 --> 01:10:41.110
science that we can't expect everyone to on top of learning their

858
01:10:41.110 --> 01:10:46.260
discipline also to become not just so so english speakers but like really

859
01:10:46.260 --> 01:10:51.140
high level english speakers who can be all right academic articles an it's

860
01:10:51.140 --> 01:10:56.130
does not reasonable so to make signs more inclusive it basically

861
01:10:56.130 --> 01:11:01.450
turns lot of people off people who might be really good scientists but you don't want to be conference

862
01:11:01.450 --> 01:11:06.290
presenting an english ah you know having to read everything all the time in english having to pay

863
01:11:06.290 --> 01:11:11.330
for their work to be translated in english having to have everything take three times as

864
01:11:11.330 --> 01:11:16.480
long because the doing at another lank tried like it's not and inclusive model and so

865
01:11:16.480 --> 01:11:21.410
if we can get away from english or any of them in its happens

866
01:11:21.410 --> 01:11:26.470
to be english but lingua franca model for science and towards multi mingle model

867
01:11:26.470 --> 01:11:31.700
and when say finance mean like research don't mean i'm pure

868
01:11:31.700 --> 01:11:36.470
and natural sciences mean all types of research

869
01:11:36.470 --> 01:11:40.910
does not answer question felucca went off in kind of ranch there bostock frighted

870
01:11:40.910 --> 01:11:44.700
by the eunice go on declaration now

871
01:11:44.700 --> 01:11:49.700
don't has time to wrap up no doubt or hope

872
01:11:49.700 --> 01:11:54.850
and kate thank you so much again for the invitation to speak and for the great stout participation

873
01:11:54.850 --> 01:11:58.230
in interaction added was lovely thank you ma'am