CMU Advanced NLP 2024 (9) Experimental Design and Human Annotation/transcript.srt

1
00:00:00,719 --> 00:00:07,480
so to get started I want to show an

2
00:00:04,120 --> 00:00:10,320
example of the scientific method I took

3
00:00:07,480 --> 00:00:12,920
this directly from Wikipedia but it's

4
00:00:10,320 --> 00:00:15,320
actually uh pretty nice it's a pretty

5
00:00:12,920 --> 00:00:17,480
nice and concise summary of what we

6
00:00:15,320 --> 00:00:19,439
should do when we're coming up with new

7
00:00:17,480 --> 00:00:22,160
uh kind of research

8
00:00:19,439 --> 00:00:24,039
projects and we start with an

9
00:00:22,160 --> 00:00:26,840
observation or question we do research

10
00:00:24,039 --> 00:00:28,599
of the topic area we form a hypothesis

11
00:00:26,840 --> 00:00:31,439
we test it with an experiment analyze

12
00:00:28,599 --> 00:00:33,600
data and Report conclusions

13
00:00:31,439 --> 00:00:35,640
and even if we're doing kind of an

14
00:00:33,600 --> 00:00:37,480
engineering based project still this

15
00:00:35,640 --> 00:00:42,079
thinking of the stuff that we're doing

16
00:00:37,480 --> 00:00:44,399
in a framework like this can help you a

17
00:00:42,079 --> 00:00:46,079
lot so uh the first thing I'd like to

18
00:00:44,399 --> 00:00:49,120
talk about is identifying good research

19
00:00:46,079 --> 00:00:51,800
directions and so I'm going to look at

20
00:00:49,120 --> 00:00:53,640
that from the observation question

21
00:00:51,800 --> 00:00:56,320
perspective

22
00:00:53,640 --> 00:00:58,480
here so if we think about why we do

23
00:00:56,320 --> 00:01:01,160
research uh particularly why we do

24
00:00:58,480 --> 00:01:04,199
research on natural language process in

25
00:01:01,160 --> 00:01:07,159
um there's a couple reasons why the

26
00:01:04,199 --> 00:01:09,439
first is application driven research and

27
00:01:07,159 --> 00:01:13,159
usually this is I would like to make a

28
00:01:09,439 --> 00:01:15,040
useful system or make one work better so

29
00:01:13,159 --> 00:01:18,479
uh you know this is probably the great

30
00:01:15,040 --> 00:01:20,280
majority of NLP research then separately

31
00:01:18,479 --> 00:01:21,960
from that there's curiosity driven

32
00:01:20,280 --> 00:01:24,560
research which is like I would like to

33
00:01:21,960 --> 00:01:27,360
know more about language or the world

34
00:01:24,560 --> 00:01:29,159
viewed through language and so this

35
00:01:27,360 --> 00:01:31,840
doesn't necessarily have to be

36
00:01:29,159 --> 00:01:31,840
immediately

37
00:01:32,000 --> 00:01:37,280
like a downstream application that users

38
00:01:35,399 --> 00:01:39,159
are using will immediately get better

39
00:01:37,280 --> 00:01:40,439
it's more like we have a burning

40
00:01:39,159 --> 00:01:43,159
question that we would like to answer

41
00:01:40,439 --> 00:01:47,399
and we want to answer

42
00:01:43,159 --> 00:01:48,640
it so NLP encompasses both uh sometimes

43
00:01:47,399 --> 00:01:50,479
if you read a paper you'll have

44
00:01:48,640 --> 00:01:54,360
something that's doing both uh

45
00:01:50,479 --> 00:01:56,439
especially like analyzing the internals

46
00:01:54,360 --> 00:01:58,079
or training dynamics of a a neural

47
00:01:56,439 --> 00:01:59,920
network to answer a curiosity-driven

48
00:01:58,079 --> 00:02:02,439
question and then applying that to come

49
00:01:59,920 --> 00:02:04,840
up with a better method that makes work

50
00:02:02,439 --> 00:02:06,560
better I I would like to say though that

51
00:02:04,840 --> 00:02:09,119
it's kind of rare that there's a paper

52
00:02:06,560 --> 00:02:10,879
that does both of them really well uh

53
00:02:09,119 --> 00:02:13,160
and so usually one of them is kind of

54
00:02:10,879 --> 00:02:14,599
the main focus and I think you can be

55
00:02:13,160 --> 00:02:17,680
well served by choosing which one is

56
00:02:14,599 --> 00:02:20,560
your main focus and then kind of uh the

57
00:02:17,680 --> 00:02:23,560
other might come as a additional uh

58
00:02:20,560 --> 00:02:23,560
bonus on top of

59
00:02:23,920 --> 00:02:28,760
that so here are a few examples of

60
00:02:27,160 --> 00:02:32,800
application driven

61
00:02:28,760 --> 00:02:35,239
research so for example pay at all uh

62
00:02:32,800 --> 00:02:37,840
they proposed the task of sentiment

63
00:02:35,239 --> 00:02:39,879
analysis um so actually there was a

64
00:02:37,840 --> 00:02:41,879
paper 22 years ago that proposed the

65
00:02:39,879 --> 00:02:44,879
task of sentiment analysis it might seem

66
00:02:41,879 --> 00:02:46,760
very you know normal nowadays but uh

67
00:02:44,879 --> 00:02:49,519
there was a paper that proposed it back

68
00:02:46,760 --> 00:02:52,840
then and they proposed sentiment

69
00:02:49,519 --> 00:02:54,200
analysis because um labeling articles

70
00:02:52,840 --> 00:02:57,480
with their sentiment would provide

71
00:02:54,200 --> 00:02:59,760
succinct summaries to the readers um so

72
00:02:57,480 --> 00:03:03,319
they basically wanted to provide

73
00:02:59,760 --> 00:03:03,319
information to readers and that would be

74
00:03:03,400 --> 00:03:09,000
useful another paper by ready at all

75
00:03:06,440 --> 00:03:11,519
2019 proposes a task of conversational

76
00:03:09,000 --> 00:03:13,640
question answering uh because an

77
00:03:11,519 --> 00:03:15,599
inability to build and maintain common

78
00:03:13,640 --> 00:03:17,680
ground is part of the reason why virtual

79
00:03:15,599 --> 00:03:20,159
assistant usually don't seem like

80
00:03:17,680 --> 00:03:22,040
competent conversational Partners so

81
00:03:20,159 --> 00:03:24,519
when you're talking to your Alexa or

82
00:03:22,040 --> 00:03:27,000
your Google uh home or something like

83
00:03:24,519 --> 00:03:28,599
this you might ask it a question and

84
00:03:27,000 --> 00:03:30,120
then after you asked it a question you

85
00:03:28,599 --> 00:03:31,480
ask it another question but it doesn't

86
00:03:30,120 --> 00:03:32,879
go back to the contexts that you had

87
00:03:31,480 --> 00:03:34,519
before and they wanted to solve this

88
00:03:32,879 --> 00:03:36,040
problem so they proposed this data set

89
00:03:34,519 --> 00:03:40,000
for

90
00:03:36,040 --> 00:03:41,720
it um Gerel propos a method for bottom

91
00:03:40,000 --> 00:03:43,159
up abstractive summarization because

92
00:03:41,720 --> 00:03:44,760
neural network-based methods for

93
00:03:43,159 --> 00:03:46,879
abstractive summarization produce

94
00:03:44,760 --> 00:03:49,000
outputs that are fluent but perform

95
00:03:46,879 --> 00:03:51,120
poorly a Content selection so they had a

96
00:03:49,000 --> 00:03:53,000
problem they had a task already in mind

97
00:03:51,120 --> 00:03:54,239
they weren't proposing a new task and

98
00:03:53,000 --> 00:03:56,040
they there was a problem with the

99
00:03:54,239 --> 00:03:58,760
existing system so they fixed

100
00:03:56,040 --> 00:04:00,400
it and then Kudo and Richardson proposed

101
00:03:58,760 --> 00:04:02,920
a method for un supervised word

102
00:04:00,400 --> 00:04:04,799
segmentation namely sentence piece uh

103
00:04:02,920 --> 00:04:06,439
because language dependent processing

104
00:04:04,799 --> 00:04:08,920
makes it hard to train multilingual

105
00:04:06,439 --> 00:04:10,360
models as we have to carefully manage

106
00:04:08,920 --> 00:04:12,720
the configurations of pre- and

107
00:04:10,360 --> 00:04:15,879
post-processors per language so they

108
00:04:12,720 --> 00:04:17,519
tried to make things easier uh so like

109
00:04:15,879 --> 00:04:19,600
you can see all of these things like the

110
00:04:17,519 --> 00:04:21,919
first two are proposing new tasks to

111
00:04:19,600 --> 00:04:23,880
solve and they're doing it from the

112
00:04:21,919 --> 00:04:25,919
point of view of uh creating something

113
00:04:23,880 --> 00:04:29,120
useful for users the second two are

114
00:04:25,919 --> 00:04:30,440
proposing new methods the first one is

115
00:04:29,120 --> 00:04:34,360
like improving

116
00:04:30,440 --> 00:04:36,320
accuracy um so it's this is the most

117
00:04:34,360 --> 00:04:37,639
common most commonly people say I have a

118
00:04:36,320 --> 00:04:39,120
test that I want to solve there's a

119
00:04:37,639 --> 00:04:41,280
problem with accuracy I want to improve

120
00:04:39,120 --> 00:04:43,960
it but you can also improve other things

121
00:04:41,280 --> 00:04:45,880
so you can improve like convenience or

122
00:04:43,960 --> 00:04:47,320
uh you can Pro improve efficiency or

123
00:04:45,880 --> 00:04:51,720
other things like that so all of those

124
00:04:47,320 --> 00:04:51,720
are you know perfectly reasonable

125
00:04:52,120 --> 00:04:57,320
things I also have some examples of

126
00:04:54,639 --> 00:04:59,120
curiosity driven research these are

127
00:04:57,320 --> 00:05:00,360
actually harder to find in the ACL

128
00:04:59,120 --> 00:05:03,120
anthology

129
00:05:00,360 --> 00:05:06,400
it's definitely the minority case but

130
00:05:03,120 --> 00:05:09,160
they still do exist um so for example

131
00:05:06,400 --> 00:05:10,960
rank at all 2017 asked what is the

132
00:05:09,160 --> 00:05:13,800
difference between the language of real

133
00:05:10,960 --> 00:05:17,000
news with that of satire hoaxes and

134
00:05:13,800 --> 00:05:18,800
propaganda so they were not attempting

135
00:05:17,000 --> 00:05:21,039
to create a system for fake news

136
00:05:18,800 --> 00:05:23,199
detection that was not their goal here

137
00:05:21,039 --> 00:05:24,600
their go their goal was just to figure

138
00:05:23,199 --> 00:05:26,240
out what were the different linguistic

139
00:05:24,600 --> 00:05:28,000
characteristics and they found that

140
00:05:26,240 --> 00:05:29,720
scientifically interesting maybe

141
00:05:28,000 --> 00:05:31,280
Downstream that would be useful but that

142
00:05:29,720 --> 00:05:35,080
wasn't the point of their

143
00:05:31,280 --> 00:05:36,960
paper another one uh curell at all ask

144
00:05:35,080 --> 00:05:38,960
are all languages equally hard to

145
00:05:36,960 --> 00:05:41,000
language model and so basically they

146
00:05:38,960 --> 00:05:42,440
wanted to know are all languages just

147
00:05:41,000 --> 00:05:45,520
character strings and so language

148
00:05:42,440 --> 00:05:47,479
modeling them is uh similarly easy or

149
00:05:45,520 --> 00:05:49,120
are there certain characteristics of

150
00:05:47,479 --> 00:05:51,080
language that make them easier or harder

151
00:05:49,120 --> 00:05:54,000
to model with the current architectures

152
00:05:51,080 --> 00:05:55,520
that we have um and so they didn't

153
00:05:54,000 --> 00:05:57,039
propose a new architecture they didn't

154
00:05:55,520 --> 00:06:00,479
propose to improve anything they just

155
00:05:57,039 --> 00:06:02,400
proposed to examine this question

156
00:06:00,479 --> 00:06:04,280
um and also Tenny at all this is

157
00:06:02,400 --> 00:06:06,880
actually an extremely impactful work

158
00:06:04,280 --> 00:06:09,319
Downstream but uh they weren't improving

159
00:06:06,880 --> 00:06:11,520
anything they just Quantified where

160
00:06:09,319 --> 00:06:14,440
specific types of linguistic information

161
00:06:11,520 --> 00:06:16,720
are encoded in birs so they found that

162
00:06:14,440 --> 00:06:18,840
for example syntax was encoded better in

163
00:06:16,720 --> 00:06:20,560
the early layers semantics in the later

164
00:06:18,840 --> 00:06:22,520
layers and then if you go further you

165
00:06:20,560 --> 00:06:25,280
you have other fine grain things like

166
00:06:22,520 --> 00:06:27,599
pragne style

167
00:06:25,280 --> 00:06:30,400
information so I I think you can kind of

168
00:06:27,599 --> 00:06:32,120
see the difference between these two um

169
00:06:30,400 --> 00:06:34,800
are there any questions

170
00:06:32,120 --> 00:06:40,199
about

171
00:06:34,800 --> 00:06:41,720
this no okay let's be that so the next

172
00:06:40,199 --> 00:06:43,680
question which I think a lot of people

173
00:06:41,720 --> 00:06:46,240
might be asking particularly with

174
00:06:43,680 --> 00:06:47,720
respect to assignment 4 which requires

175
00:06:46,240 --> 00:06:51,039
you to come up with something novel to

176
00:06:47,720 --> 00:06:53,240
do is how do we uh get research

177
00:06:51,039 --> 00:06:57,360
ideas

178
00:06:53,240 --> 00:07:02,280
and the way we can do this is uh twofold

179
00:06:57,360 --> 00:07:04,479
so um one is kind of we want to turn a

180
00:07:02,280 --> 00:07:07,120
concrete understanding of existing

181
00:07:04,479 --> 00:07:10,120
research's failings into a higher level

182
00:07:07,120 --> 00:07:12,560
experimental question and the two ways

183
00:07:10,120 --> 00:07:15,240
that I normally characterize doing this

184
00:07:12,560 --> 00:07:19,319
are bottom up discovery of research

185
00:07:15,240 --> 00:07:21,080
ideas um or the way the way I

186
00:07:19,319 --> 00:07:24,479
characterize this is bottom up discovery

187
00:07:21,080 --> 00:07:27,000
of research ideas and this is a great

188
00:07:24,479 --> 00:07:29,120
tool for making incremental progress on

189
00:07:27,000 --> 00:07:32,039
existing systems on tasks that we really

190
00:07:29,120 --> 00:07:35,400
care about or expanding the scope of a

191
00:07:32,039 --> 00:07:37,680
task that we care about so uh some

192
00:07:35,400 --> 00:07:41,879
examples of this would be like in

193
00:07:37,680 --> 00:07:45,639
assignment number three you uh look

194
00:07:41,879 --> 00:07:47,720
let's say you're looking at

195
00:07:45,639 --> 00:07:50,159
um let's say you're looking at the

196
00:07:47,720 --> 00:07:53,840
question answering performance

197
00:07:50,159 --> 00:07:58,280
of models of multilingual models on

198
00:07:53,840 --> 00:08:01,479
different languages um and you for

199
00:07:58,280 --> 00:08:03,159
assignment three you implement a couple

200
00:08:01,479 --> 00:08:05,240
multilingual models on different

201
00:08:03,159 --> 00:08:06,560
languages you run them you look at the

202
00:08:05,240 --> 00:08:08,400
results and you identify that

203
00:08:06,560 --> 00:08:10,080
multilingual models are particularly bad

204
00:08:08,400 --> 00:08:12,919
at answering questions about named

205
00:08:10,080 --> 00:08:14,680
entities and so now you have looked at

206
00:08:12,919 --> 00:08:17,759
the output you have decided that that's

207
00:08:14,680 --> 00:08:20,199
a big problem um you can go in and

208
00:08:17,759 --> 00:08:22,080
improve it so this is a great tool for

209
00:08:20,199 --> 00:08:23,720
incremental progress and like in fact

210
00:08:22,080 --> 00:08:26,520
doing this really effectively has been

211
00:08:23,720 --> 00:08:31,000
very effective in my own research career

212
00:08:26,520 --> 00:08:34,680
like we uh if I feel like I I like to

213
00:08:31,000 --> 00:08:36,279
look at data I try to do that a lot and

214
00:08:34,680 --> 00:08:38,440
by doing that I identify the most

215
00:08:36,279 --> 00:08:40,200
frequent problems and because of that

216
00:08:38,440 --> 00:08:42,039
when I fix those problems my accuracy

217
00:08:40,200 --> 00:08:44,560
goes up a lot more than people who pick

218
00:08:42,039 --> 00:08:46,880
the less good problems right and so if

219
00:08:44,560 --> 00:08:49,440
we want our accuracy to go up uh I'm

220
00:08:46,880 --> 00:08:51,360
more efficient at you know improving

221
00:08:49,440 --> 00:08:53,240
things on the other hand there's

222
00:08:51,360 --> 00:08:55,399
something uh from the opposite direction

223
00:08:53,240 --> 00:08:57,080
is moving from a higher level question

224
00:08:55,399 --> 00:08:57,800
to a lower level concrete testing of

225
00:08:57,080 --> 00:09:00,120
that

226
00:08:57,800 --> 00:09:01,760
question um so this could be tap down

227
00:09:00,120 --> 00:09:02,760
Design This is tap down design of

228
00:09:01,760 --> 00:09:06,360
research

229
00:09:02,760 --> 00:09:08,399
ideas this favors bigger ideas but these

230
00:09:06,360 --> 00:09:10,240
ideas can be disconnected from reality

231
00:09:08,399 --> 00:09:13,880
or they could be not solving the right

232
00:09:10,240 --> 00:09:17,079
problems so the typical like very very

233
00:09:13,880 --> 00:09:18,800
successful example of this is um neural

234
00:09:17,079 --> 00:09:20,800
machine translation or something like

235
00:09:18,800 --> 00:09:22,720
this neural machine translations neural

236
00:09:20,800 --> 00:09:26,399
sequence sequence

237
00:09:22,720 --> 00:09:30,040
models this came out of a few people

238
00:09:26,399 --> 00:09:32,040
like Jeff Hinton and yua

239
00:09:30,040 --> 00:09:33,480
believing for a very long time that

240
00:09:32,040 --> 00:09:35,760
neural networks were the right way to

241
00:09:33,480 --> 00:09:37,800
solve lots of problems uh despite the

242
00:09:35,760 --> 00:09:39,640
fact that there wasn't like super

243
00:09:37,800 --> 00:09:42,279
concrete evidence of that for a long

244
00:09:39,640 --> 00:09:43,399
time and so they had this idea which was

245
00:09:42,279 --> 00:09:47,399
like we should be doing things with

246
00:09:43,399 --> 00:09:49,440
neural networks and uh they you know

247
00:09:47,399 --> 00:09:50,720
they successfully executed that and now

248
00:09:49,440 --> 00:09:52,200
everybody is doing things with neural

249
00:09:50,720 --> 00:09:56,560
networks so they made a really huge

250
00:09:52,200 --> 00:09:58,160
revolution in the research space um that

251
00:09:56,560 --> 00:09:59,720
that's great that's a great example of a

252
00:09:58,160 --> 00:10:02,839
successful topown IDE IDE but the

253
00:09:59,720 --> 00:10:05,519
problem is uh for every example like

254
00:10:02,839 --> 00:10:07,560
that there's a thousand uh top down

255
00:10:05,519 --> 00:10:10,760
ideas in the graveyard of not being very

256
00:10:07,560 --> 00:10:12,600
you know effective so I I think um in

257
00:10:10,760 --> 00:10:14,519
order to do something like this you

258
00:10:12,600 --> 00:10:16,200
better have a very strong conviction or

259
00:10:14,519 --> 00:10:18,079
you better have maybe some initial

260
00:10:16,200 --> 00:10:20,920
evidence or a very strong intuition

261
00:10:18,079 --> 00:10:22,320
about why this might be a good idea and

262
00:10:20,920 --> 00:10:25,240
uh you would be able to test that

263
00:10:22,320 --> 00:10:27,240
intuition through intermediate steps uh

264
00:10:25,240 --> 00:10:31,040
to to demonstrate like through toy data

265
00:10:27,240 --> 00:10:31,040
or other stuff like that

266
00:10:31,720 --> 00:10:38,360
um cool so these are kind of the general

267
00:10:36,360 --> 00:10:40,839
ways that we can come up with research

268
00:10:38,360 --> 00:10:42,519
ideas the next thing that we want to do

269
00:10:40,839 --> 00:10:44,480
is research our topic area were there

270
00:10:42,519 --> 00:10:46,720
any questions about bottom up versus top

271
00:10:44,480 --> 00:10:49,120
down I'm going to talk about effective

272
00:10:46,720 --> 00:10:51,920
strategies to bottom up stuff in uh in

273
00:10:49,120 --> 00:10:54,360
two weeks uh so we can talk more about

274
00:10:51,920 --> 00:10:56,800
that then

275
00:10:54,360 --> 00:11:00,959
but okay if not I'll move

276
00:10:56,800 --> 00:11:05,079
on so next uh we have research topic

277
00:11:00,959 --> 00:11:07,360
areas so this is about how you will do

278
00:11:05,079 --> 00:11:10,320
assignment three which is researching uh

279
00:11:07,360 --> 00:11:13,240
topic area getting forming a very good

280
00:11:10,320 --> 00:11:15,680
understanding of the topic that you're

281
00:11:13,240 --> 00:11:18,800
trying to handle and so there's a bunch

282
00:11:15,680 --> 00:11:22,800
of different ways you can do this uh the

283
00:11:18,800 --> 00:11:25,680
first one is keyword search and so you

284
00:11:22,800 --> 00:11:27,839
look something up on Google Scholar or

285
00:11:25,680 --> 00:11:29,480
something uh finding older and newer

286
00:11:27,839 --> 00:11:32,880
papers so this is like following the

287
00:11:29,480 --> 00:11:35,360
tracks of papers you can uh read the

288
00:11:32,880 --> 00:11:39,160
abstract and intro uh read the details

289
00:11:35,360 --> 00:11:43,760
of most relevant papers and I don't do

290
00:11:39,160 --> 00:11:45,440
this as much now but um when I was a

291
00:11:43,760 --> 00:11:47,360
graduate student I would often make a

292
00:11:45,440 --> 00:11:49,800
short summary of the paper to make sure

293
00:11:47,360 --> 00:11:54,680
I really understood the details uh

294
00:11:49,800 --> 00:11:56,000
because also now I teach a class um and

295
00:11:54,680 --> 00:11:58,240
actually making these slides is very

296
00:11:56,000 --> 00:12:00,120
useful for me so going back into the

297
00:11:58,240 --> 00:12:03,440
Transformer slide slides you know that

298
00:12:00,120 --> 00:12:05,160
kind of serves as my um you know my way

299
00:12:03,440 --> 00:12:06,800
of digesting papers and making sure that

300
00:12:05,160 --> 00:12:08,160
I can explain them and if you're not

301
00:12:06,800 --> 00:12:10,480
teaching a class and you can go in and

302
00:12:08,160 --> 00:12:13,560
make a summary into it yourselves so

303
00:12:10,480 --> 00:12:16,480
that can confirm uh solidify your memory

304
00:12:13,560 --> 00:12:19,360
and like confirm your uh ability to

305
00:12:16,480 --> 00:12:19,360
understand everything that's in

306
00:12:20,639 --> 00:12:27,120
there cool um so next I'd like to talk

307
00:12:23,639 --> 00:12:29,600
about some sources of papers in NLP um

308
00:12:27,120 --> 00:12:31,800
one really good source uh is the ACL

309
00:12:29,600 --> 00:12:33,720
Anthology another good source is Google

310
00:12:31,800 --> 00:12:36,120
Scholar um they both have their

311
00:12:33,720 --> 00:12:37,959
advantages and their disadvantages um

312
00:12:36,120 --> 00:12:39,800
increasingly actually I realized now

313
00:12:37,959 --> 00:12:41,959
that I should add this to my slides but

314
00:12:39,800 --> 00:12:43,639
increasingly a lot of good uh papers in

315
00:12:41,959 --> 00:12:47,120
NLP are also published in machine

316
00:12:43,639 --> 00:12:51,199
learning conferences so like icml or NPS

317
00:12:47,120 --> 00:12:53,040
or um uh I clear or things like that the

318
00:12:51,199 --> 00:12:54,920
problem is the ACL Anthology is way

319
00:12:53,040 --> 00:12:56,600
better than any of them at like

320
00:12:54,920 --> 00:13:00,360
organizing the papers in an easy to

321
00:12:56,600 --> 00:13:03,560
process way so I I think um I I'll talk

322
00:13:00,360 --> 00:13:06,000
about this uh for now and so the ACL

323
00:13:03,560 --> 00:13:08,800
Anthology covers many uh prestigious

324
00:13:06,000 --> 00:13:11,639
venues in NLP it has all of these ones

325
00:13:08,800 --> 00:13:15,160
here this figure is a little bit old uh

326
00:13:11,639 --> 00:13:18,839
I I made it in 21 2021 but you know it

327
00:13:15,160 --> 00:13:22,959
reaches up to the present day and what I

328
00:13:18,839 --> 00:13:25,880
do often is I can start with the past 3

329
00:13:22,959 --> 00:13:30,160
to 5 years of several top venues in here

330
00:13:25,880 --> 00:13:33,880
like ACL emnlp uh nackle and tackle and

331
00:13:30,160 --> 00:13:36,360
go in and do uh keyword search and so

332
00:13:33,880 --> 00:13:36,360
like let's

333
00:13:38,760 --> 00:13:43,600
say let's say I was interested in

334
00:13:44,639 --> 00:13:49,519
multilingual multilingual large language

335
00:13:47,600 --> 00:13:52,079
models and evaluating them or some way

336
00:13:49,519 --> 00:13:54,279
so I would go to ACL and then I would

337
00:13:52,079 --> 00:13:57,560
just put in multi

338
00:13:54,279 --> 00:14:01,360
lingual um and you get a wonderful paper

339
00:13:57,560 --> 00:14:01,360
by by some research are

340
00:14:01,480 --> 00:14:06,440
named that was not intentional I didn't

341
00:14:03,639 --> 00:14:08,800
know that was going to happen but um so

342
00:14:06,440 --> 00:14:11,240
on the Fly crosslingual masking for

343
00:14:08,800 --> 00:14:12,959
multilingual pre-training um scaling

344
00:14:11,240 --> 00:14:15,040
multilingual corpora and language models

345
00:14:12,959 --> 00:14:18,120
to 500 languages that seems pretty

346
00:14:15,040 --> 00:14:19,880
pretty relevant evaluating multilingual

347
00:14:18,120 --> 00:14:22,000
compositional generalization so you can

348
00:14:19,880 --> 00:14:27,680
just go through here and see a bunch of

349
00:14:22,000 --> 00:14:30,680
papers that like um that could be

350
00:14:27,680 --> 00:14:30,680
useful

351
00:14:32,240 --> 00:14:35,199
and you could uh if you're doing a more

352
00:14:33,800 --> 00:14:36,920
machine learning oriented thing you can

353
00:14:35,199 --> 00:14:38,920
do the same thing for like the nurs

354
00:14:36,920 --> 00:14:41,480
proceedings or the icml proceedings or

355
00:14:38,920 --> 00:14:41,480
something like

356
00:14:41,800 --> 00:14:48,120
that um separately from this you can go

357
00:14:44,839 --> 00:14:50,920
through Google Scholar um this allows

358
00:14:48,120 --> 00:14:52,560
for a search of papers by keyword and so

359
00:14:50,920 --> 00:14:54,440
if I write like neural entity

360
00:14:52,560 --> 00:14:56,360
recognition it will give neural

361
00:14:54,440 --> 00:15:00,040
architectures for identity recognition

362
00:14:56,360 --> 00:15:03,399
all of these things like this um you can

363
00:15:00,040 --> 00:15:06,800
view the more recent papers so like for

364
00:15:03,399 --> 00:15:10,120
example uh if you're researching uh kind

365
00:15:06,800 --> 00:15:12,759
of generic topic that a lot of people

366
00:15:10,120 --> 00:15:14,639
use uh a lot of people do research on

367
00:15:12,759 --> 00:15:18,399
you might be getting papers from like

368
00:15:14,639 --> 00:15:19,920
1998 or something like this and you know

369
00:15:18,399 --> 00:15:21,639
they might be useful but honestly the

370
00:15:19,920 --> 00:15:23,519
methodology has changed so much since

371
00:15:21,639 --> 00:15:24,680
then that most methodical papers from

372
00:15:23,519 --> 00:15:26,959
that long ago are probably not going to

373
00:15:24,680 --> 00:15:29,480
be very useful um so you can view the

374
00:15:26,959 --> 00:15:31,079
recent papers another really useful

375
00:15:29,480 --> 00:15:33,759
thing that you can do is view papers

376
00:15:31,079 --> 00:15:35,319
that site the current paper and you can

377
00:15:33,759 --> 00:15:39,560
even click on this and then you can

378
00:15:35,319 --> 00:15:42,519
search within the sighting papers so

379
00:15:39,560 --> 00:15:44,399
um like let's say I want to know about

380
00:15:42,519 --> 00:15:45,620
how

381
00:15:44,399 --> 00:15:48,730
people

382
00:15:45,620 --> 00:15:48,730
[Music]

383
00:15:50,720 --> 00:15:55,720
do let's say I want to see if anybody

384
00:15:53,199 --> 00:15:59,639
does neural entity recognition with uh

385
00:15:55,720 --> 00:16:02,160
State space models so I do like stage

386
00:15:59,639 --> 00:16:05,399
space

387
00:16:02,160 --> 00:16:09,040
model and then I search within the

388
00:16:05,399 --> 00:16:12,279
citing articles and I'm able to find

389
00:16:09,040 --> 00:16:14,319
three articles that at least cite this

390
00:16:12,279 --> 00:16:17,759
paper and and talk about State space

391
00:16:14,319 --> 00:16:20,319
models so

392
00:16:17,759 --> 00:16:21,600
um none of these seem particularly

393
00:16:20,319 --> 00:16:23,240
relevant to what I was looking for but

394
00:16:21,600 --> 00:16:26,800
you get the idea like this can be a

395
00:16:23,240 --> 00:16:26,800
useful tool for finding more recent

396
00:16:27,519 --> 00:16:30,519
things

397
00:16:33,639 --> 00:16:40,480
and then finding older papers this is

398
00:16:36,279 --> 00:16:42,839
also relatively easy um so you read the

399
00:16:40,480 --> 00:16:44,319
papers that you're interested in and

400
00:16:42,839 --> 00:16:45,480
then it will have back blinks to older

401
00:16:44,319 --> 00:16:47,519
papers and you look them up in the

402
00:16:45,480 --> 00:16:50,000
references this is how I I find older

403
00:16:47,519 --> 00:16:53,600
papers that might be

404
00:16:50,000 --> 00:16:57,800
relevant um and so the these are the

405
00:16:53,600 --> 00:16:59,720
tools that I use um some other so I I'd

406
00:16:57,800 --> 00:17:03,600
like to give a few caveats about Google

407
00:16:59,720 --> 00:17:06,120
Scholar and uh things like Twitter or

408
00:17:03,600 --> 00:17:08,360
LinkedIn or something like this they

409
00:17:06,120 --> 00:17:10,720
give you very biased views on all the

410
00:17:08,360 --> 00:17:14,600
papers that are out there um because

411
00:17:10,720 --> 00:17:16,919
they sort for popularity basically so um

412
00:17:14,600 --> 00:17:19,439
actually if you're looking at like

413
00:17:16,919 --> 00:17:22,000
Twitter or LinkedIn or something like

414
00:17:19,439 --> 00:17:23,679
that you can actually get a pretty bleak

415
00:17:22,000 --> 00:17:25,360
view on natural language processing and

416
00:17:23,679 --> 00:17:28,000
say all anybody is doing is training

417
00:17:25,360 --> 00:17:30,080
large language models because you know

418
00:17:28,000 --> 00:17:31,720
these things tend to become you know

419
00:17:30,080 --> 00:17:33,520
popular and then they get Amplified by

420
00:17:31,720 --> 00:17:35,840
algorithms and stuff like that when in

421
00:17:33,520 --> 00:17:37,440
fact like the landscape is much richer

422
00:17:35,840 --> 00:17:40,400
which is why I do definitely suggest

423
00:17:37,440 --> 00:17:42,000
that you like actually look through uh

424
00:17:40,400 --> 00:17:43,880
conference proceedings and stuff and

425
00:17:42,000 --> 00:17:46,720
find papers that are not you know

426
00:17:43,880 --> 00:17:48,520
Amplified as much so um I I definitely

427
00:17:46,720 --> 00:17:50,840
highly recommend doing this in addition

428
00:17:48,520 --> 00:17:52,480
to you know Google Scholar or social

429
00:17:50,840 --> 00:17:54,640
media or other things like that that

430
00:17:52,480 --> 00:17:54,640
might

431
00:17:56,600 --> 00:18:01,760
be cool um I'd also like to mention a

432
00:18:00,200 --> 00:18:04,000
thing about the ups and downs of

433
00:18:01,760 --> 00:18:07,559
preemptive surveys

434
00:18:04,000 --> 00:18:10,440
so um surveying extensively before doing

435
00:18:07,559 --> 00:18:12,840
research uh has a bunch of good sides so

436
00:18:10,440 --> 00:18:14,000
it prevents you from duplicating work so

437
00:18:12,840 --> 00:18:15,039
somebody else might have done a very

438
00:18:14,000 --> 00:18:18,080
similar

439
00:18:15,039 --> 00:18:20,480
thing um it also increases your toolbox

440
00:18:18,080 --> 00:18:21,600
of methods so you know if it's a problem

441
00:18:20,480 --> 00:18:25,400
that a lot of people have worked on

442
00:18:21,600 --> 00:18:27,120
before then you know it helps uh give

443
00:18:25,400 --> 00:18:30,320
you ideas of methods that you could be

444
00:18:27,120 --> 00:18:35,600
using um however in a way it also kind

445
00:18:30,320 --> 00:18:38,720
of constrains your thinking so um if you

446
00:18:35,600 --> 00:18:42,480
like on once you have built up a very

447
00:18:38,720 --> 00:18:45,440
extensive survey of like ways to do

448
00:18:42,480 --> 00:18:47,240
things you tend to like move away from

449
00:18:45,440 --> 00:18:48,799
there when in fact like if you thought

450
00:18:47,240 --> 00:18:50,080
just thought of ways to solve problems

451
00:18:48,799 --> 00:18:52,360
without looking at everything you might

452
00:18:50,080 --> 00:18:54,799
come up with something over here might

453
00:18:52,360 --> 00:18:56,400
actually be a good idea right um and so

454
00:18:54,799 --> 00:18:58,600
there's this really nice essay it was

455
00:18:56,400 --> 00:19:00,799
actually shared uh shared with me by

456
00:18:58,600 --> 00:19:02,440
Chris Manning from Sanford um it's

457
00:19:00,799 --> 00:19:04,720
called how to build an economics model

458
00:19:02,440 --> 00:19:06,679
in your spare time it's about it's from

459
00:19:04,720 --> 00:19:08,880
a Nobel Prize winner in economics but

460
00:19:06,679 --> 00:19:10,480
he's talking about how when he tries to

461
00:19:08,880 --> 00:19:13,039
come up with new and like important

462
00:19:10,480 --> 00:19:15,840
ideas he doesn't look at economics

463
00:19:13,039 --> 00:19:19,679
journals he looks at the newspaper and

464
00:19:15,840 --> 00:19:21,919
tries to uh you know

465
00:19:19,679 --> 00:19:23,480
like look at problems that people are

466
00:19:21,919 --> 00:19:24,840
talking about in the newspaper and think

467
00:19:23,480 --> 00:19:27,159
about whether there's an economic

468
00:19:24,840 --> 00:19:29,919
solution to them and so if we think

469
00:19:27,159 --> 00:19:32,880
about the anal of how we can do this in

470
00:19:29,919 --> 00:19:35,600
natural language processing you know

471
00:19:32,880 --> 00:19:37,360
maybe you don't necessarily right away

472
00:19:35,600 --> 00:19:38,799
want to do a really extensive survey

473
00:19:37,360 --> 00:19:41,080
first you might just think about like

474
00:19:38,799 --> 00:19:44,080
what's bothering you like when you're

475
00:19:41,080 --> 00:19:46,799
using chat GPT what is really

476
00:19:44,080 --> 00:19:49,600
frustrating to you uh about how it gives

477
00:19:46,799 --> 00:19:51,280
responses or um what are the things you

478
00:19:49,600 --> 00:19:53,159
wish it were possible to do through

479
00:19:51,280 --> 00:19:56,240
natural language processing but not are

480
00:19:53,159 --> 00:19:57,640
not possible to do and um then you can

481
00:19:56,240 --> 00:20:00,679
start from there you can look at you

482
00:19:57,640 --> 00:20:03,440
know what companies are doing in their

483
00:20:00,679 --> 00:20:05,799
Tech demos uh because the tech demos

484
00:20:03,440 --> 00:20:08,640
might be nice but they almost never work

485
00:20:05,799 --> 00:20:11,240
as well as the tech demo makes them seem

486
00:20:08,640 --> 00:20:13,840
like they work so that could be another

487
00:20:11,240 --> 00:20:15,720
place to get ideas um or you can look at

488
00:20:13,840 --> 00:20:17,039
papers in a related field like machine

489
00:20:15,720 --> 00:20:18,760
learning like let's say you're a machine

490
00:20:17,039 --> 00:20:21,280
learning oriented person and you really

491
00:20:18,760 --> 00:20:23,000
love like math and stuff like that it's

492
00:20:21,280 --> 00:20:25,799
like well there's this good mathematical

493
00:20:23,000 --> 00:20:27,760
tool that I think could be applicable to

494
00:20:25,799 --> 00:20:30,440
um a certain problem in NLP or something

495
00:20:27,760 --> 00:20:31,960
like that so you could do that too um

496
00:20:30,440 --> 00:20:33,960
the the final one you know comes with

497
00:20:31,960 --> 00:20:35,799
all the caveats of doing topown research

498
00:20:33,960 --> 00:20:37,320
of course so you know you need to make

499
00:20:35,799 --> 00:20:39,799
sure that that really is the correct

500
00:20:37,320 --> 00:20:42,159
tool for whatever you want to sell but

501
00:20:39,799 --> 00:20:45,280
um definitely this is something to think

502
00:20:42,159 --> 00:20:48,240
about um however for assignment three

503
00:20:45,280 --> 00:20:49,559
you need to do a survey so I'm I'm

504
00:20:48,240 --> 00:20:50,720
forcing you to do a survey for

505
00:20:49,559 --> 00:20:52,200
assignment three so if you're going to

506
00:20:50,720 --> 00:20:53,640
do something like this you can do it

507
00:20:52,200 --> 00:20:56,600
before assignment 3 and start thinking

508
00:20:53,640 --> 00:21:00,000
about what you want to be doing so um

509
00:20:56,600 --> 00:21:01,520
that's something

510
00:21:00,000 --> 00:21:03,200
uh any questions or discussion about

511
00:21:01,520 --> 00:21:06,799
that

512
00:21:03,200 --> 00:21:07,840
part this is hard I'm I'm happy to uh

513
00:21:06,799 --> 00:21:11,120
happy to

514
00:21:07,840 --> 00:21:14,039
discuss either now or in office hours or

515
00:21:11,120 --> 00:21:14,039
anything like this

516
00:21:14,200 --> 00:21:19,720
but Okay

517
00:21:17,080 --> 00:21:24,279
cool so the next thing is a for

518
00:21:19,720 --> 00:21:25,640
hypothesis so uh once you have done you

519
00:21:24,279 --> 00:21:28,600
have a general idea of what you want to

520
00:21:25,640 --> 00:21:31,240
do um and you have done a survey related

521
00:21:28,600 --> 00:21:32,480
work you can devise a final research

522
00:21:31,240 --> 00:21:34,159
question or

523
00:21:32,480 --> 00:21:37,760
hypothesis

524
00:21:34,159 --> 00:21:40,039
and so a research question is one or

525
00:21:37,760 --> 00:21:43,400
several explicit questions regarding the

526
00:21:40,039 --> 00:21:45,919
thing that you want to know um

527
00:21:43,400 --> 00:21:47,400
and this is actually pretty hard for

528
00:21:45,919 --> 00:21:49,080
people like I ask people to write

529
00:21:47,400 --> 00:21:50,880
research questions and very often they

530
00:21:49,080 --> 00:21:53,080
don't write research questions in this

531
00:21:50,880 --> 00:21:57,720
format and I have to ask people to try

532
00:21:53,080 --> 00:21:59,919
to change them and what they what I

533
00:21:57,720 --> 00:22:03,159
think they in general should be are yes

534
00:21:59,919 --> 00:22:08,120
no questions so

535
00:22:03,159 --> 00:22:10,400
it um yes no questions and you have a

536
00:22:08,120 --> 00:22:13,120
hypothesis uh about what you think the

537
00:22:10,400 --> 00:22:14,600
answer to the question may be a priori

538
00:22:13,120 --> 00:22:17,520
and that hypothesis should be

539
00:22:14,600 --> 00:22:19,919
falsifiable so basically it's if you get

540
00:22:17,520 --> 00:22:21,240
a certain result you can demonstrate

541
00:22:19,919 --> 00:22:23,120
that the answer to this question is

542
00:22:21,240 --> 00:22:24,679
probably yes if you get a different

543
00:22:23,120 --> 00:22:27,520
result you can demonstrate that the

544
00:22:24,679 --> 00:22:29,640
answer to the question is probably no

545
00:22:27,520 --> 00:22:32,400
and just to make this a little bit more

546
00:22:29,640 --> 00:22:34,360
concrete I can give a few curiosity

547
00:22:32,400 --> 00:22:36,880
driven questions and

548
00:22:34,360 --> 00:22:40,720
hypothesis C the Curiosity driven

549
00:22:36,880 --> 00:22:43,480
questions are a little bit easier so um

550
00:22:40,720 --> 00:22:45,600
we have the Curiosity driven question of

551
00:22:43,480 --> 00:22:49,679
are all language models are all

552
00:22:45,600 --> 00:22:53,559
languages equally hard to language model

553
00:22:49,679 --> 00:22:55,400
and they say uh it is unlikely that all

554
00:22:53,559 --> 00:22:56,760
languages are equally easy or that

555
00:22:55,400 --> 00:22:58,799
methods are equally good at all

556
00:22:56,760 --> 00:23:01,159
languages um so so that's their

557
00:22:58,799 --> 00:23:04,120
hypothesis so they think a priori that

558
00:23:01,159 --> 00:23:05,919
that's the case um but that might be

559
00:23:04,120 --> 00:23:08,400
falsified by getting a very strong

560
00:23:05,919 --> 00:23:10,679
result that says like no matter which

561
00:23:08,400 --> 00:23:13,760
language you're modeling many models

562
00:23:10,679 --> 00:23:18,120
that we use get get similar results

563
00:23:13,760 --> 00:23:20,400
on um what makes a particular podcast

564
00:23:18,120 --> 00:23:21,320
broadly engaging so this was an analysis

565
00:23:20,400 --> 00:23:24,400
of

566
00:23:21,320 --> 00:23:27,960
podcasts uh where they compared popular

567
00:23:24,400 --> 00:23:29,720
podcasts and unpopular podcasts or

568
00:23:27,960 --> 00:23:32,400
engaging and unengaging

569
00:23:29,720 --> 00:23:34,400
podcasts and it says uh tips such as

570
00:23:32,400 --> 00:23:37,039
reducing filler words and disfluencies

571
00:23:34,400 --> 00:23:38,840
or incorporating emotion are things that

572
00:23:37,039 --> 00:23:41,400
people had anecdotally written on the

573
00:23:38,840 --> 00:23:43,039
internet as tips to make a good podcast

574
00:23:41,400 --> 00:23:45,760
but nobody had actually empirically

575
00:23:43,039 --> 00:23:48,440
valid validated that so they wanted to

576
00:23:45,760 --> 00:23:50,000
like actually go invalidate that so they

577
00:23:48,440 --> 00:23:51,679
came up with hypotheses and they could

578
00:23:50,000 --> 00:23:55,720
demonstrate that those had good or bad

579
00:23:51,679 --> 00:23:55,720
correlation podcast being judged as

580
00:23:56,880 --> 00:24:03,600
engaging application driven questions

581
00:23:59,039 --> 00:24:03,600
and hypotheses are a little bit harder

582
00:24:04,520 --> 00:24:10,480
so here is an

583
00:24:07,640 --> 00:24:13,039
example this is an example from a paper

584
00:24:10,480 --> 00:24:18,720
that I wrote previously which

585
00:24:13,039 --> 00:24:22,080
was where and why or how and why do

586
00:24:18,720 --> 00:24:22,960
pre-trained word embeddings help neural

587
00:24:22,080 --> 00:24:25,080
machine

588
00:24:22,960 --> 00:24:26,760
translation and this was back when

589
00:24:25,080 --> 00:24:28,279
pre-training was mostly like word

590
00:24:26,760 --> 00:24:31,880
embeddings we weren't preing the whole

591
00:24:28,279 --> 00:24:34,480
body of the neural net so

592
00:24:31,880 --> 00:24:36,640
now the answers to this question are a

593
00:24:34,480 --> 00:24:37,919
little bit different but basically the

594
00:24:36,640 --> 00:24:40,080
questions that we asked is is the

595
00:24:37,919 --> 00:24:42,360
behavior of pre-training affected by

596
00:24:40,080 --> 00:24:45,960
language families and other linguistic

597
00:24:42,360 --> 00:24:49,520
features of source and Target languages

598
00:24:45,960 --> 00:24:51,360
so uh we expected that the answer to

599
00:24:49,520 --> 00:24:53,640
this would be yes it would vary across

600
00:24:51,360 --> 00:24:54,960
them do pre-trained edings help more

601
00:24:53,640 --> 00:24:57,760
when the size of the training data is

602
00:24:54,960 --> 00:24:59,039
small we expected that this would be yes

603
00:24:57,760 --> 00:25:00,640
how much does the similarity of the

604
00:24:59,039 --> 00:25:03,720
source and Target languages affect the

605
00:25:00,640 --> 00:25:06,200
efficacy of using pre-trained edings uh

606
00:25:03,720 --> 00:25:08,399
we didn't have a hypothesis about

607
00:25:06,200 --> 00:25:10,600
whether it would or not and is it

608
00:25:08,399 --> 00:25:12,320
helpful to align the embedding spaces

609
00:25:10,600 --> 00:25:14,520
between the source and Target languages

610
00:25:12,320 --> 00:25:16,039
we assume this would be yes and do

611
00:25:14,520 --> 00:25:17,640
pre-trained edings help more in

612
00:25:16,039 --> 00:25:19,360
multilingual systems as compared to

613
00:25:17,640 --> 00:25:22,679
bilingual systems and we didn't have a

614
00:25:19,360 --> 00:25:26,279
good hypothesis about that

615
00:25:22,679 --> 00:25:29,559
I another one is although recent stud uh

616
00:25:26,279 --> 00:25:32,760
sorry the question of whether and how

617
00:25:29,559 --> 00:25:35,039
contextual information benefits endtoend

618
00:25:32,760 --> 00:25:38,960
speech translation has received little

619
00:25:35,039 --> 00:25:42,480
attention and so their guess was that it

620
00:25:38,960 --> 00:25:44,880
probably would help so application

621
00:25:42,480 --> 00:25:47,120
oriented questions are a little bit

622
00:25:44,880 --> 00:25:49,200
tricky because the obvious one is like

623
00:25:47,120 --> 00:25:52,200
does X make y

624
00:25:49,200 --> 00:25:54,080
better and so you you have a method you

625
00:25:52,200 --> 00:25:55,559
think it's going to make the output

626
00:25:54,080 --> 00:25:58,120
better and so that's kind of your

627
00:25:55,559 --> 00:26:00,000
obvious research question but the

628
00:25:58,120 --> 00:26:02,080
problem is the above question or

629
00:26:00,000 --> 00:26:04,279
hypothesis is natural but it's very

630
00:26:02,080 --> 00:26:06,679
indirect so normally you also have a

631
00:26:04,279 --> 00:26:09,760
hypothesis about like why it will help

632
00:26:06,679 --> 00:26:13,279
or something like this and so if the

633
00:26:09,760 --> 00:26:15,440
answer is no after your experiments why

634
00:26:13,279 --> 00:26:18,080
is the answer

635
00:26:15,440 --> 00:26:20,640
no it could be that your original

636
00:26:18,080 --> 00:26:23,720
assumption about why a particular method

637
00:26:20,640 --> 00:26:25,039
would help was wrong which is the worst

638
00:26:23,720 --> 00:26:28,360
case scenario but you also could just

639
00:26:25,039 --> 00:26:30,559
have a bug in your code or uh your

640
00:26:28,360 --> 00:26:32,000
data set your test set might not be

641
00:26:30,559 --> 00:26:34,279
large enough so you wouldn't be able to

642
00:26:32,000 --> 00:26:35,840
get a statistically significant result

643
00:26:34,279 --> 00:26:40,039
based on the amount that it helped you

644
00:26:35,840 --> 00:26:42,960
improve or other things like that so

645
00:26:40,039 --> 00:26:44,960
what I like to do in this case is try to

646
00:26:42,960 --> 00:26:48,399
come up with the intuition about why X

647
00:26:44,960 --> 00:26:50,360
will make y better and can you think of

648
00:26:48,399 --> 00:26:52,080
other research questions or hypotheses

649
00:26:50,360 --> 00:26:54,240
that confirm or falsified these

650
00:26:52,080 --> 00:26:56,640
assumptions

651
00:26:54,240 --> 00:26:59,559
so uh some things that you can do are

652
00:26:56,640 --> 00:27:01,240
come up with like toy data or come up

653
00:26:59,559 --> 00:27:03,840
with a subset of the data where you

654
00:27:01,240 --> 00:27:06,600
think this might be correct so just to

655
00:27:03,840 --> 00:27:09,279
give an example let's say we have a

656
00:27:06,600 --> 00:27:12,159
translation model and we have a

657
00:27:09,279 --> 00:27:14,279
hypothesis that improving entity

658
00:27:12,159 --> 00:27:16,520
translation and low resource languages

659
00:27:14,279 --> 00:27:18,799
will improve translation accuracy and we

660
00:27:16,520 --> 00:27:21,399
run an experiment or actually maybe this

661
00:27:18,799 --> 00:27:23,760
is an even better one we we have a

662
00:27:21,399 --> 00:27:26,240
hypothesis that incorporating contextual

663
00:27:23,760 --> 00:27:28,799
information in speech translation will

664
00:27:26,240 --> 00:27:31,760
help translation results

665
00:27:28,799 --> 00:27:36,480
so incorporating context in machine

666
00:27:31,760 --> 00:27:37,600
translation has been a very old topic

667
00:27:36,480 --> 00:27:41,279
like people have been trying to do this

668
00:27:37,600 --> 00:27:43,559
for a very long time but for a long time

669
00:27:41,279 --> 00:27:45,200
the conclusion was that it essentially

670
00:27:43,559 --> 00:27:46,519
wasn't helping translation people would

671
00:27:45,200 --> 00:27:48,039
incorporate contacts through neural

672
00:27:46,519 --> 00:27:50,960
networks or other things like that and

673
00:27:48,039 --> 00:27:53,320
it just wasn't improving the results

674
00:27:50,960 --> 00:27:55,320
significantly and in the end the reason

675
00:27:53,320 --> 00:27:57,960
why was because there just weren't

676
00:27:55,320 --> 00:27:59,799
enough examples where contextual

677
00:27:57,960 --> 00:28:02,200
information was useful in the data sets

678
00:27:59,799 --> 00:28:06,360
that everybody was using so people were

679
00:28:02,200 --> 00:28:09,080
using really long news sentences to try

680
00:28:06,360 --> 00:28:10,880
to figure out where uh whether context

681
00:28:09,080 --> 00:28:12,440
was helping but really long new

682
00:28:10,880 --> 00:28:14,000
sentences have so much information

683
00:28:12,440 --> 00:28:16,080
included in them that you can mostly

684
00:28:14,000 --> 00:28:20,120
translate sentence by sentence and get

685
00:28:16,080 --> 00:28:21,880
it right like 95% of the time so the

686
00:28:20,120 --> 00:28:23,600
problem wasn't that any of the methods

687
00:28:21,880 --> 00:28:26,799
that people were proposing were bad it

688
00:28:23,600 --> 00:28:29,559
was just that they weren't effective

689
00:28:26,799 --> 00:28:31,440
enough to see big enough uh results and

690
00:28:29,559 --> 00:28:33,159
so then people Chang the data set to

691
00:28:31,440 --> 00:28:34,720
like conversations or something like

692
00:28:33,159 --> 00:28:37,399
that and in conversations they're very

693
00:28:34,720 --> 00:28:39,159
contextual yeah very short utterances

694
00:28:37,399 --> 00:28:41,440
and once you started doing things like

695
00:28:39,159 --> 00:28:45,840
that then the same methods like exactly

696
00:28:41,440 --> 00:28:48,640
the same methods were um were helping

697
00:28:45,840 --> 00:28:51,120
when they weren't helping before and

698
00:28:48,640 --> 00:28:52,720
so the underlying assumption about

699
00:28:51,120 --> 00:28:56,240
incorporating context information is

700
00:28:52,720 --> 00:28:58,159
that context will be helpful and or

701
00:28:56,240 --> 00:29:01,760
context is necessary

702
00:28:58,159 --> 00:29:03,880
to you know do translation well so does

703
00:29:01,760 --> 00:29:06,880
anyone have an idea about how you could

704
00:29:03,880 --> 00:29:06,880
like actually verify that

705
00:29:10,880 --> 00:29:16,519
assumption any idea yeah simplest way

706
00:29:14,000 --> 00:29:19,120
would be just give an El way to set and

707
00:29:16,519 --> 00:29:21,000
then have a measure of okay if it in

708
00:29:19,120 --> 00:29:23,679
more than

709
00:29:21,000 --> 00:29:25,519
x% um and how would that verify the

710
00:29:23,679 --> 00:29:28,480
assumption that context is

711
00:29:25,519 --> 00:29:30,720
necessary so we're asking a question

712
00:29:28,480 --> 00:29:33,480
whether context is helpful in the proect

713
00:29:30,720 --> 00:29:36,000
you're doing that uh we're asking

714
00:29:33,480 --> 00:29:39,240
whether

715
00:29:36,000 --> 00:29:40,840
so we're asking kind of a a two-part the

716
00:29:39,240 --> 00:29:44,080
main question is whether context is

717
00:29:40,840 --> 00:29:45,559
helpful given a particular you know

718
00:29:44,080 --> 00:29:47,240
experimental setup right so like

719
00:29:45,559 --> 00:29:50,440
training data

720
00:29:47,240 --> 00:29:52,039
set modeling method and training

721
00:29:50,440 --> 00:29:54,679
algorithm and evaluation algorithm

722
00:29:52,039 --> 00:29:56,480
that's kind of the big final result that

723
00:29:54,679 --> 00:29:58,840
you want to get in your paper but

724
00:29:56,480 --> 00:30:01,399
there's kind of a the question which is

725
00:29:58,840 --> 00:30:04,360
is context even necessary to translate

726
00:30:01,399 --> 00:30:06,559
well you train a model with context and

727
00:30:04,360 --> 00:30:08,200
one without context you train a model

728
00:30:06,559 --> 00:30:10,679
with context and one without context but

729
00:30:08,200 --> 00:30:14,080
what if your model of context is really

730
00:30:10,679 --> 00:30:15,399
bad J the same model you have the same

731
00:30:14,080 --> 00:30:16,840
model architecture but let's say your

732
00:30:15,399 --> 00:30:18,559
model architecture is really bad at

733
00:30:16,840 --> 00:30:19,919
capturing context so then maybe it's a

734
00:30:18,559 --> 00:30:22,399
problem of your model architecture and

735
00:30:19,919 --> 00:30:24,720
context is necessary or helpful but your

736
00:30:22,399 --> 00:30:27,399
model just isn't very good at capture

737
00:30:24,720 --> 00:30:29,720
human yeah exactly so this is one thing

738
00:30:27,399 --> 00:30:31,960
that people can do so there was a

739
00:30:29,720 --> 00:30:34,240
interesting paper um let me see if I can

740
00:30:31,960 --> 00:30:34,240
find

741
00:30:39,960 --> 00:30:49,080
it so this is a paper from a long time

742
00:30:45,760 --> 00:30:51,600
ago where they did something like

743
00:30:49,080 --> 00:30:53,360
this um it's evaluating machine

744
00:30:51,600 --> 00:30:54,480
translation systems with second language

745
00:30:53,360 --> 00:30:57,399
proficiency

746
00:30:54,480 --> 00:31:01,240
tests and basically what they did is

747
00:30:57,399 --> 00:31:03,519
they had these English proficiency tests

748
00:31:01,240 --> 00:31:05,320
for uh I think it was like middle

749
00:31:03,519 --> 00:31:07,480
schoolers or high schoolers or something

750
00:31:05,320 --> 00:31:09,600
like this and then they used machine

751
00:31:07,480 --> 00:31:11,240
translation systems to translate them

752
00:31:09,600 --> 00:31:13,600
into Japanese and then they asked

753
00:31:11,240 --> 00:31:19,720
Japanese students to solve them in

754
00:31:13,600 --> 00:31:19,720
japanies and so what they did is they

755
00:31:20,000 --> 00:31:26,159
asked uh Anonymous system G and

756
00:31:23,679 --> 00:31:28,200
Anonymous system Y which are Google and

757
00:31:26,159 --> 00:31:32,360
Yahoo

758
00:31:28,200 --> 00:31:34,720
and uh and a human without context and a

759
00:31:32,360 --> 00:31:36,279
human with context to translate them so

760
00:31:34,720 --> 00:31:38,720
they ask humans to translate each

761
00:31:36,279 --> 00:31:40,880
sentence without giving any context and

762
00:31:38,720 --> 00:31:44,320
they ask humans to translate each uh

763
00:31:40,880 --> 00:31:46,399
sentence with giving context and what

764
00:31:44,320 --> 00:31:48,960
they were able to find was in this case

765
00:31:46,399 --> 00:31:50,080
humans with context the Japanese

766
00:31:48,960 --> 00:31:53,080
students were able to answer the

767
00:31:50,080 --> 00:31:55,360
questions most of the time um whereas if

768
00:31:53,080 --> 00:31:57,559
they translated without contexts like G

769
00:31:55,360 --> 00:31:59,039
and Y were doing at that time actually

770
00:31:57,559 --> 00:32:01,320
why was almost as good as human

771
00:31:59,039 --> 00:32:04,080
translators at you know achieving the

772
00:32:01,320 --> 00:32:05,440
the task so but basically like the

773
00:32:04,080 --> 00:32:09,159
important thing here is they were able

774
00:32:05,440 --> 00:32:11,039
to confirm their you know idea that in

775
00:32:09,159 --> 00:32:12,519
this case humans with context were much

776
00:32:11,039 --> 00:32:13,799
better than humans without context so

777
00:32:12,519 --> 00:32:16,279
that would verify your like sub

778
00:32:13,799 --> 00:32:18,080
assumption right and so this is just

779
00:32:16,279 --> 00:32:20,279
like one

780
00:32:18,080 --> 00:32:22,240
example this is just one example of

781
00:32:20,279 --> 00:32:25,960
something that you can

782
00:32:22,240 --> 00:32:27,480
do uh but the basic idea is like your

783
00:32:25,960 --> 00:32:29,320
final result is that you want build of

784
00:32:27,480 --> 00:32:30,799
system that does better on some

785
00:32:29,320 --> 00:32:32,159
Benchmark that you care about there's a

786
00:32:30,799 --> 00:32:33,600
bunch of things that go into whether it

787
00:32:32,159 --> 00:32:36,159
does better or not your evaluation

788
00:32:33,600 --> 00:32:38,960
system your model your training data

789
00:32:36,159 --> 00:32:41,559
your training your evaluation data set

790
00:32:38,960 --> 00:32:43,080
um and things like that so can you break

791
00:32:41,559 --> 00:32:45,360
that down into sub questions that you

792
00:32:43,080 --> 00:32:48,039
could ask where you could verify that

793
00:32:45,360 --> 00:32:49,720
it's working or not uh based on whether

794
00:32:48,039 --> 00:32:51,600
those things are happening another thing

795
00:32:49,720 --> 00:32:53,159
people do an ml oriented things is

796
00:32:51,600 --> 00:32:54,919
create a toy data set where they know

797
00:32:53,159 --> 00:32:57,200
the phenomenon they're interested in

798
00:32:54,919 --> 00:32:59,679
exists and train their models on there

799
00:32:57,200 --> 00:33:02,919
and make sure that they work there um so

800
00:32:59,679 --> 00:33:02,919
that's another thing that you can take

801
00:33:03,120 --> 00:33:07,639
that cool um any questions about

802
00:33:08,080 --> 00:33:12,760
this okay

803
00:33:10,200 --> 00:33:16,519
s so the next thing is running

804
00:33:12,760 --> 00:33:19,000
experiments um so in order to do this

805
00:33:16,519 --> 00:33:21,399
you'll find data that will answer your

806
00:33:19,000 --> 00:33:23,639
research question uh run experiments and

807
00:33:21,399 --> 00:33:25,720
calculate numbers uh calculate

808
00:33:23,639 --> 00:33:28,279
significant differences and analyze

809
00:33:25,720 --> 00:33:31,080
effects whoops

810
00:33:28,279 --> 00:33:35,519
and so this is a basic pipeline that we

811
00:33:31,080 --> 00:33:37,760
want to follow so obtaining test data so

812
00:33:35,519 --> 00:33:41,200
in order to obtain test data uh we would

813
00:33:37,760 --> 00:33:42,799
like to find data sets um so if you're

814
00:33:41,200 --> 00:33:46,200
building on previous work the safest

815
00:33:42,799 --> 00:33:48,960
thing that you can do um is start with

816
00:33:46,200 --> 00:33:51,919
the same data sets if you're answering a

817
00:33:48,960 --> 00:33:53,799
new question um you can think about can

818
00:33:51,919 --> 00:33:55,399
you repurpose other data sets to answer

819
00:33:53,799 --> 00:33:57,679
the question so very often there will be

820
00:33:55,399 --> 00:34:00,080
a data set that is uh appropriate for

821
00:33:57,679 --> 00:34:03,360
answer answering your question um and

822
00:34:00,080 --> 00:34:05,760
you can go and find that um actually our

823
00:34:03,360 --> 00:34:06,919
our wonderful TJ has created a system

824
00:34:05,760 --> 00:34:08,800
called datafinder that will

825
00:34:06,919 --> 00:34:11,159
automatically find it for you so if you

826
00:34:08,800 --> 00:34:13,679
want to uh search for data sets you can

827
00:34:11,159 --> 00:34:16,760
use his system or ask him about it but

828
00:34:13,679 --> 00:34:20,359
um uh but if no appropriate data set

829
00:34:16,760 --> 00:34:24,359
exists you can uh create your own and

830
00:34:20,359 --> 00:34:25,879
particularly for industry use cases it's

831
00:34:24,359 --> 00:34:28,119
very common that you need to go in and

832
00:34:25,879 --> 00:34:30,040
create your own or if you're planning on

833
00:34:28,119 --> 00:34:31,639
doing research in Academia afterwards

834
00:34:30,040 --> 00:34:33,119
very often you'll come up with a

835
00:34:31,639 --> 00:34:34,639
research question where no data set

836
00:34:33,119 --> 00:34:36,679
exists so you'll have to create your own

837
00:34:34,639 --> 00:34:38,960
anyway so this is something that's

838
00:34:36,679 --> 00:34:41,639
really important to be able to do well

839
00:34:38,960 --> 00:34:44,639
uh in most

840
00:34:41,639 --> 00:34:49,240
cases um so I'll be talking about how to

841
00:34:44,639 --> 00:34:53,280
do all of these so data set lists um the

842
00:34:49,240 --> 00:34:55,159
best one I think by far in uh natural

843
00:34:53,280 --> 00:34:58,359
language processing nowadays is hugging

844
00:34:55,159 --> 00:35:02,960
face data sets um there's also other

845
00:34:58,359 --> 00:35:05,359
data resources like um elra is uh

846
00:35:02,960 --> 00:35:07,240
another one kind of by the more

847
00:35:05,359 --> 00:35:09,800
traditional natural language processing

848
00:35:07,240 --> 00:35:12,960
Community there's also the LDC the

849
00:35:09,800 --> 00:35:15,680
linguistic data uh Consortium and there

850
00:35:12,960 --> 00:35:17,119
are some older heavily annotated data

851
00:35:15,680 --> 00:35:20,040
sets that are only available through

852
00:35:17,119 --> 00:35:22,000
those at CMU you have the ability to

853
00:35:20,040 --> 00:35:24,520
download things from LDC so if you find

854
00:35:22,000 --> 00:35:26,960
an LDC data set in any papers that

855
00:35:24,520 --> 00:35:29,640
you're doing or online um you need

856
00:35:26,960 --> 00:35:31,000
register for that and I I'm the person

857
00:35:29,640 --> 00:35:33,280
who's in charge of it so I'll give you

858
00:35:31,000 --> 00:35:35,520
access and then uh and then you can use

859
00:35:33,280 --> 00:35:37,400
it um there's also things like papers

860
00:35:35,520 --> 00:35:39,680
with code and papers with code basically

861
00:35:37,400 --> 00:35:41,359
automatically extracts uh kind of like

862
00:35:39,680 --> 00:35:42,839
the names of data sets so even some

863
00:35:41,359 --> 00:35:45,599
things that don't appear on a hug and

864
00:35:42,839 --> 00:35:45,599
place will appear

865
00:35:46,359 --> 00:35:52,440
there so annotating data um when you

866
00:35:50,640 --> 00:35:54,599
annotate data you first need to decide

867
00:35:52,440 --> 00:35:57,599
how much to annotate sample appropriate

868
00:35:54,599 --> 00:36:00,240
data create annotation guidelines

869
00:35:57,599 --> 00:36:03,160
uh either annotate yourself or hire and

870
00:36:00,240 --> 00:36:05,839
supervis annotators and evaluate

871
00:36:03,160 --> 00:36:07,720
quality so a very common problem that a

872
00:36:05,839 --> 00:36:10,240
lot of people ask me is how much test

873
00:36:07,720 --> 00:36:12,800
data do you need

874
00:36:10,240 --> 00:36:14,800
and I'm going to talk about uh

875
00:36:12,800 --> 00:36:17,520
statistical significance tests in a

876
00:36:14,800 --> 00:36:19,520
second but um basically you need to have

877
00:36:17,520 --> 00:36:23,240
enough to have a statistically

878
00:36:19,520 --> 00:36:28,119
significant difference um between

879
00:36:23,240 --> 00:36:32,079
methods and the way you do this actually

880
00:36:28,119 --> 00:36:32,079
sorry very quickly let me

881
00:36:33,240 --> 00:36:37,599
check I rearrange my slides and I want

882
00:36:35,560 --> 00:36:40,359
to make sure that I didn't accidentally

883
00:36:37,599 --> 00:36:42,280
um I didn't accidentally remove the

884
00:36:40,359 --> 00:36:44,520
slides on statistical significance which

885
00:36:42,280 --> 00:36:44,520
would be

886
00:36:51,680 --> 00:36:57,880
a okay

887
00:36:55,240 --> 00:36:59,200
um sorry hang on one second I just

888
00:36:57,880 --> 00:37:02,240
realized that I don't have the slides

889
00:36:59,200 --> 00:37:03,839
for a statistical significance on this

890
00:37:02,240 --> 00:37:05,280
presentation so let me grab them from

891
00:37:03,839 --> 00:37:09,440
the

892
00:37:05,280 --> 00:37:09,440
last uh the last

893
00:37:10,520 --> 00:37:14,640
us this is is pretty

894
00:37:25,599 --> 00:37:28,599
important

895
00:37:33,160 --> 00:37:38,599
okay so yeah let me explain statistical

896
00:37:35,560 --> 00:37:40,319
significance here um so basically when

897
00:37:38,599 --> 00:37:43,319
we're doing statistical

898
00:37:40,319 --> 00:37:44,680
testing um let's say we have two models

899
00:37:43,319 --> 00:37:47,800
with similar

900
00:37:44,680 --> 00:37:50,160
accuracies and these models with similar

901
00:37:47,800 --> 00:37:52,240
accuracies let's say model one is a

902
00:37:50,160 --> 00:37:56,880
generative model model two is a

903
00:37:52,240 --> 00:37:58,520
discriminative model and we say uh data

904
00:37:56,880 --> 00:38:00,200
set one we have this result on data set

905
00:37:58,520 --> 00:38:02,480
two we have another result on data set

906
00:38:00,200 --> 00:38:04,720
three we have uh another

907
00:38:02,480 --> 00:38:06,440
result and so then the question is how

908
00:38:04,720 --> 00:38:09,480
can we tell if the differences are due

909
00:38:06,440 --> 00:38:13,839
to consistent trends that uh will hold

910
00:38:09,480 --> 00:38:16,119
on other data sets or um if they are

911
00:38:13,839 --> 00:38:18,480
kind of random noise due to the fact

912
00:38:16,119 --> 00:38:21,000
that we have one

913
00:38:18,480 --> 00:38:24,200
uh due to the fact that you know data

914
00:38:21,000 --> 00:38:25,640
sets vary models vary um and so the way

915
00:38:24,200 --> 00:38:28,319
we do this is through statistical

916
00:38:25,640 --> 00:38:31,839
significance testing

917
00:38:28,319 --> 00:38:34,319
um so I'm going to cover this briefly in

918
00:38:31,839 --> 00:38:36,920
this class but you can see a drawer at

919
00:38:34,319 --> 00:38:38,640
all for an overview and also we're going

920
00:38:36,920 --> 00:38:41,520
to have a recitation on how to actually

921
00:38:38,640 --> 00:38:44,280
run statistical significance tests so um

922
00:38:41,520 --> 00:38:47,920
you can take a look at that

923
00:38:44,280 --> 00:38:51,680
there and so the basic idea is given a

924
00:38:47,920 --> 00:38:54,280
quantity we test um certain values of

925
00:38:51,680 --> 00:38:57,880
uncertainty with respect to the quantity

926
00:38:54,280 --> 00:38:59,960
so number one is a p value and the P

927
00:38:57,880 --> 00:39:02,240
value is what is the probability that a

928
00:38:59,960 --> 00:39:06,119
difference with another quantity is by

929
00:39:02,240 --> 00:39:08,359
chance and so a lower uh P value means

930
00:39:06,119 --> 00:39:11,839
more likelihood of having a significant

931
00:39:08,359 --> 00:39:13,200
difference usually the threshold for

932
00:39:11,839 --> 00:39:16,520
saying that we have a significant

933
00:39:13,200 --> 00:39:20,280
difference is there's a 5% chance

934
00:39:16,520 --> 00:39:22,160
0.05 that this difference between the

935
00:39:20,280 --> 00:39:25,760
models was due to chance or like data

936
00:39:22,160 --> 00:39:28,520
sampling or things like that uh so p uh

937
00:39:25,760 --> 00:39:30,880
less than 0.05 is kind of a threshold

938
00:39:28,520 --> 00:39:30,880
for

939
00:39:31,119 --> 00:39:35,680
significance another thing that we can

940
00:39:33,040 --> 00:39:38,720
measure is confidence intervals and the

941
00:39:35,680 --> 00:39:40,760
confidence interval is um what is the

942
00:39:38,720 --> 00:39:42,560
range under which we could expect

943
00:39:40,760 --> 00:39:44,760
another trial to fall and I'll talk

944
00:39:42,560 --> 00:39:47,359
about both of

945
00:39:44,760 --> 00:39:49,280
these um there's another concept called

946
00:39:47,359 --> 00:39:53,880
paired versus unpaired

947
00:39:49,280 --> 00:39:56,680
tests and in unpaired test comp this

948
00:39:53,880 --> 00:39:59,480
means um we compare the means of a

949
00:39:56,680 --> 00:40:02,359
quantity on two unrelated

950
00:39:59,480 --> 00:40:04,040
groups so an example could be the test

951
00:40:02,359 --> 00:40:07,040
of the significance of a difference of

952
00:40:04,040 --> 00:40:09,160
accuracies of a model on two data sets

953
00:40:07,040 --> 00:40:12,400
so like let's say I have data set number

954
00:40:09,160 --> 00:40:16,440
one and data set number two what is the

955
00:40:12,400 --> 00:40:18,000
likelihood that the um there's actually

956
00:40:16,440 --> 00:40:20,839
a real difference in the data sets as

957
00:40:18,000 --> 00:40:23,400
opposed to just random uh random

958
00:40:20,839 --> 00:40:26,599
sampling RS between

959
00:40:23,400 --> 00:40:28,560
them in contrast AED test compares the

960
00:40:26,599 --> 00:40:31,400
means of a quantity on one data set

961
00:40:28,560 --> 00:40:32,480
under two conditions and so an example

962
00:40:31,400 --> 00:40:33,760
of this could be testing the

963
00:40:32,480 --> 00:40:37,319
significance of a difference of

964
00:40:33,760 --> 00:40:39,640
accuracies of two models on one data set

965
00:40:37,319 --> 00:40:42,000
so this is a really important difference

966
00:40:39,640 --> 00:40:43,960
and the reason why it's a really

967
00:40:42,000 --> 00:40:45,520
important difference well number one

968
00:40:43,960 --> 00:40:49,119
we're most commonly interested in the

969
00:40:45,520 --> 00:40:51,839
letter number two if we can make

970
00:40:49,119 --> 00:40:54,280
assumptions about

971
00:40:51,839 --> 00:40:56,079
the association of the points in the

972
00:40:54,280 --> 00:40:58,680
data set we're much much more likely to

973
00:40:56,079 --> 00:41:00,440
get a significant result because we can

974
00:40:58,680 --> 00:41:02,240
um we can look at the difference of the

975
00:41:00,440 --> 00:41:06,000
models on individual data points as

976
00:41:02,240 --> 00:41:10,400
opposed to um uh as opposed to looking

977
00:41:06,000 --> 00:41:10,400
at just the difference in the

978
00:41:10,520 --> 00:41:16,839
means so one example of a statistical

979
00:41:13,760 --> 00:41:18,280
significance test is a bootstrap test

980
00:41:16,839 --> 00:41:19,760
and the bootstrap test is really

981
00:41:18,280 --> 00:41:21,680
convenient because you can implement it

982
00:41:19,760 --> 00:41:25,160
for any evaluation metric that you want

983
00:41:21,680 --> 00:41:26,880
to be using and so in NLP we can use

984
00:41:25,160 --> 00:41:29,560
lots of different evaluations metrics we

985
00:41:26,880 --> 00:41:31,119
can use an evaluation metric like um

986
00:41:29,560 --> 00:41:34,160
accuracy but we can also use an

987
00:41:31,119 --> 00:41:37,400
evaluation metric like fmeasure for

988
00:41:34,160 --> 00:41:40,560
classification or a blue score or

989
00:41:37,400 --> 00:41:43,599
character F score or word error rate or

990
00:41:40,560 --> 00:41:48,440
something like that for um for various

991
00:41:43,599 --> 00:41:50,720
tasks and this is applicable to any any

992
00:41:48,440 --> 00:41:54,000
metric you want to use uh any quantity

993
00:41:50,720 --> 00:41:57,319
you want to measure also so the basic

994
00:41:54,000 --> 00:41:59,079
idea of a bootstrap test is a method

995
00:41:57,319 --> 00:42:02,520
that can measure P values and confidence

996
00:41:59,079 --> 00:42:06,040
intervals by resampling data and so the

997
00:42:02,520 --> 00:42:08,480
way you do this is you sample subsets

998
00:42:06,040 --> 00:42:11,960
from your death Dev test set with

999
00:42:08,480 --> 00:42:14,720
replacement so you might sample 10,000

1000
00:42:11,960 --> 00:42:19,599
times and you measure accuracy on these

1001
00:42:14,720 --> 00:42:22,520
many subsets and then you take

1002
00:42:19,599 --> 00:42:25,640
the you look at all of the accuracies

1003
00:42:22,520 --> 00:42:27,680
that you got on these subsample data

1004
00:42:25,640 --> 00:42:31,079
sets and then you take the middle

1005
00:42:27,680 --> 00:42:32,640
percentile range like 2.5 to 97.5 and

1006
00:42:31,079 --> 00:42:34,960
you can treat that as a confidence

1007
00:42:32,640 --> 00:42:37,640
interval the 95% confidence interval

1008
00:42:34,960 --> 00:42:40,720
about where you're like 95% certain that

1009
00:42:37,640 --> 00:42:40,720
your results will fall in

1010
00:42:40,880 --> 00:42:48,240
here another thing that you can do is

1011
00:42:45,119 --> 00:42:50,040
you can do a paired test and what the

1012
00:42:48,240 --> 00:42:51,200
paired test does is it measures the

1013
00:42:50,040 --> 00:42:53,359
number of

1014
00:42:51,200 --> 00:42:55,839
winds um

1015
00:42:53,359 --> 00:42:57,720
if and you measure the percentage of

1016
00:42:55,839 --> 00:43:00,920
winds and this is the confidence that a

1017
00:42:57,720 --> 00:43:03,280
gain in accuracy is not by chance um and

1018
00:43:00,920 --> 00:43:05,920
so this could be one minus the P value

1019
00:43:03,280 --> 00:43:07,960
of the paired test so this is easy to

1020
00:43:05,920 --> 00:43:09,960
implement applicable to any evaluation

1021
00:43:07,960 --> 00:43:13,480
measure but somewhat biased on small

1022
00:43:09,960 --> 00:43:17,240
data sets um just to maybe I can give a

1023
00:43:13,480 --> 00:43:19,920
more concrete example so let's say we

1024
00:43:17,240 --> 00:43:27,520
have a classification data set what you

1025
00:43:19,920 --> 00:43:30,400
can do is um let's say we have a b c d e

1026
00:43:27,520 --> 00:43:36,960
e or

1027
00:43:30,400 --> 00:43:39,559
um X1 X2 X3 X4

1028
00:43:36,960 --> 00:43:44,520
X5 so this is our our classification

1029
00:43:39,559 --> 00:43:47,440
data set and um we have system

1030
00:43:44,520 --> 00:43:52,000
one system

1031
00:43:47,440 --> 00:43:53,760
two and we have right right right right

1032
00:43:52,000 --> 00:43:56,599
wrong

1033
00:43:53,760 --> 00:44:00,440
right uh right wrong

1034
00:43:56,599 --> 00:44:03,040
long right or something like this and so

1035
00:44:00,440 --> 00:44:07,079
what we do is we randomly sample a sub

1036
00:44:03,040 --> 00:44:08,760
data set um and let's say this is like

1037
00:44:07,079 --> 00:44:10,440
X3

1038
00:44:08,760 --> 00:44:13,599
X2

1039
00:44:10,440 --> 00:44:17,599
X4 X1

1040
00:44:13,599 --> 00:44:20,440
X2 and so this is our subd data set uh

1041
00:44:17,599 --> 00:44:20,440
what we do

1042
00:44:20,640 --> 00:44:28,920
is um so X3 would be

1043
00:44:23,520 --> 00:44:34,559
01 X2 would be 1 one X4 would be one Zer

1044
00:44:28,920 --> 00:44:39,079
X X1 would be 1 one and

1045
00:44:34,559 --> 00:44:42,319
then uh X X2 would be one and so the

1046
00:44:39,079 --> 00:44:45,319
overall accuracy here

1047
00:44:42,319 --> 00:44:45,319
is

1048
00:44:45,480 --> 00:44:50,240
60% and

1049
00:44:47,440 --> 00:44:51,880
80% so if we didn't do any statistical

1050
00:44:50,240 --> 00:44:55,400
significance test we might say oh system

1051
00:44:51,880 --> 00:44:57,680
2 is better obviously um but if we do

1052
00:44:55,400 --> 00:45:01,079
the significance test this is one sample

1053
00:44:57,680 --> 00:45:03,119
from the bootstrap test in

1054
00:45:01,079 --> 00:45:07,040
here

1055
00:45:03,119 --> 00:45:09,079
now we get like 80% and 80% and it's

1056
00:45:07,040 --> 00:45:11,079
like okay actually maybe in some cases

1057
00:45:09,079 --> 00:45:13,480
these systems AR equally good maybe

1058
00:45:11,079 --> 00:45:16,079
there's a tie or if we sampled another

1059
00:45:13,480 --> 00:45:19,079
one uh let's say we

1060
00:45:16,079 --> 00:45:19,079
sampled

1061
00:45:19,359 --> 00:45:27,319
uh

1062
00:45:20,960 --> 00:45:30,680
X4 X1 X2 X4 X1

1063
00:45:27,319 --> 00:45:36,160
um um then we would get something like

1064
00:45:30,680 --> 00:45:37,559
one Z one one one one 1 0 1 one this

1065
00:45:36,160 --> 00:45:40,440
would be

1066
00:45:37,559 --> 00:45:42,559
100% And this would be

1067
00:45:40,440 --> 00:45:44,960
60% and

1068
00:45:42,559 --> 00:45:47,000
so in some cases depending on how we

1069
00:45:44,960 --> 00:45:48,440
sample actually system one wins and so

1070
00:45:47,000 --> 00:45:51,440
you count the number of times that

1071
00:45:48,440 --> 00:45:52,880
system two wins based on um based on

1072
00:45:51,440 --> 00:45:54,280
these sub samples you count the number

1073
00:45:52,880 --> 00:45:56,400
of times that system one wins and you

1074
00:45:54,280 --> 00:45:59,000
count the number of times you get a tie

1075
00:45:56,400 --> 00:46:00,920
and only in the case where system two or

1076
00:45:59,000 --> 00:46:03,680
like the better system wins more than

1077
00:46:00,920 --> 00:46:06,280
95% of the time you say that there's a

1078
00:46:03,680 --> 00:46:08,599
significant difference be these or

1079
00:46:06,280 --> 00:46:10,720
alternatively you could also look at the

1080
00:46:08,599 --> 00:46:15,960
confidence intervals by saying okay I

1081
00:46:10,720 --> 00:46:19,000
sampled um like 90 95% of the time uh

1082
00:46:15,960 --> 00:46:20,920
the accuracy of system one is uh like

1083
00:46:19,000 --> 00:46:23,640
80% or lower and so that would give you

1084
00:46:20,920 --> 00:46:23,640
the upper L

1085
00:46:23,760 --> 00:46:29,599
calculation so yeah sorry this is a very

1086
00:46:27,480 --> 00:46:31,760
uh very quick overview of this but the

1087
00:46:29,599 --> 00:46:34,240
reason why this is useful is let's say

1088
00:46:31,760 --> 00:46:36,160
you create a very small data set if you

1089
00:46:34,240 --> 00:46:38,400
create a very small data set this is

1090
00:46:36,160 --> 00:46:39,880
going to give you a very it's going to

1091
00:46:38,400 --> 00:46:41,319
be very hard to get a statistically

1092
00:46:39,880 --> 00:46:44,319
significant result on this data set

1093
00:46:41,319 --> 00:46:47,200
because it's tiny right and you know

1094
00:46:44,319 --> 00:46:50,640
quite frequently you're going to be

1095
00:46:47,200 --> 00:46:53,400
sampling um you're going to be sampling

1096
00:46:50,640 --> 00:46:55,400
data sets like this where the model like

1097
00:46:53,400 --> 00:46:56,640
where model one wins quite frequently

1098
00:46:55,400 --> 00:46:58,520
you're going to be sampling other data

1099
00:46:56,640 --> 00:47:00,359
sets where key wins and basically you're

1100
00:46:58,520 --> 00:47:02,920
not going to be able to say with

1101
00:47:00,359 --> 00:47:04,480
confidence which model is better because

1102
00:47:02,920 --> 00:47:06,359
you just don't have enough data to say

1103
00:47:04,480 --> 00:47:07,880
that but as you make your data set

1104
00:47:06,359 --> 00:47:11,119
bigger and bigger it becomes easier and

1105
00:47:07,880 --> 00:47:14,240
easier to get a significant result and

1106
00:47:11,119 --> 00:47:17,400
so uh because you're more sure that you

1107
00:47:14,240 --> 00:47:20,960
didn't just randomly pick data that

1108
00:47:17,400 --> 00:47:25,400
model two is better at

1109
00:47:20,960 --> 00:47:28,440
uh so um there's also other varieties

1110
00:47:25,400 --> 00:47:31,240
ofest there's things like T tests for

1111
00:47:28,440 --> 00:47:34,720
unpaired unpaired outputs and paired T

1112
00:47:31,240 --> 00:47:38,079
tests for paired outputs those work when

1113
00:47:34,720 --> 00:47:40,440
your um outputs are eddied so they work

1114
00:47:38,079 --> 00:47:43,599
for accuracy because the accuracy is

1115
00:47:40,440 --> 00:47:46,440
just you add all the add all the ones

1116
00:47:43,599 --> 00:47:48,680
and then divide by the um the number of

1117
00:47:46,440 --> 00:47:50,960
instances and that gives you an accuracy

1118
00:47:48,680 --> 00:47:57,880
that doesn't work for something like

1119
00:47:50,960 --> 00:48:03,599
fmeasure um because fmeasure is um 2 *

1120
00:47:57,880 --> 00:48:07,319
Precision Time recall / Precision plus

1121
00:48:03,599 --> 00:48:08,040
recall um and precision and recall uh

1122
00:48:07,319 --> 00:48:10,640
you

1123
00:48:08,040 --> 00:48:12,920
can like a T Test works for this but

1124
00:48:10,640 --> 00:48:15,160
there's a non-additive component of f

1125
00:48:12,920 --> 00:48:16,680
measure so you can't calculate

1126
00:48:15,160 --> 00:48:19,280
statistically significant differences in

1127
00:48:16,680 --> 00:48:21,079
F measure using a key test in that case

1128
00:48:19,280 --> 00:48:23,000
you're basically you have to use a

1129
00:48:21,079 --> 00:48:24,920
bootstrap method like this in order to

1130
00:48:23,000 --> 00:48:29,040
get it to work or you need to do some

1131
00:48:24,920 --> 00:48:29,040
really complex math but I I just

1132
00:48:29,760 --> 00:48:33,920
use cool um are there any questions

1133
00:48:32,680 --> 00:48:35,520
about this I guess we'll have a code

1134
00:48:33,920 --> 00:48:37,680
example in the recitation so you can go

1135
00:48:35,520 --> 00:48:39,599
in and take a look at that there's also

1136
00:48:37,680 --> 00:48:42,599
tons of code examples

1137
00:48:39,599 --> 00:48:42,599
online

1138
00:48:42,960 --> 00:48:49,440
um is that

1139
00:48:45,720 --> 00:48:52,400
okay okay sounds good um so now let me

1140
00:48:49,440 --> 00:48:54,599
uh let me go back to the actual slides

1141
00:48:52,400 --> 00:48:57,400
for

1142
00:48:54,599 --> 00:49:00,559
today and given those statist uh the

1143
00:48:57,400 --> 00:49:04,119
results about statistical signicance um

1144
00:49:00,559 --> 00:49:06,040
how can we estimate how much testing

1145
00:49:04,119 --> 00:49:07,920
data is enough and there's a method

1146
00:49:06,040 --> 00:49:11,079
called Power analysis that allows you to

1147
00:49:07,920 --> 00:49:13,359
do this and basically the idea of power

1148
00:49:11,079 --> 00:49:16,680
analysis is that you make an assumption

1149
00:49:13,359 --> 00:49:18,880
about the effect size between settings

1150
00:49:16,680 --> 00:49:20,680
um for example the expected accuracy

1151
00:49:18,880 --> 00:49:23,480
difference between tested

1152
00:49:20,680 --> 00:49:26,480
models and given the effect size a

1153
00:49:23,480 --> 00:49:28,880
significance threshold and significant

1154
00:49:26,480 --> 00:49:30,839
threshold you can determine how much

1155
00:49:28,880 --> 00:49:32,680
data is necessary to get a significant

1156
00:49:30,839 --> 00:49:36,680
effect in most

1157
00:49:32,680 --> 00:49:39,319
CLS and so to give an example

1158
00:49:36,680 --> 00:49:41,559
again let's say we're talking about the

1159
00:49:39,319 --> 00:49:45,880
accuracy let's say we have a baseline

1160
00:49:41,559 --> 00:49:49,079
model and we have a um we have a

1161
00:49:45,880 --> 00:49:52,280
baseline model and then we also have our

1162
00:49:49,079 --> 00:49:54,000
uh propos model and we know kind of from

1163
00:49:52,280 --> 00:49:55,599
experience that the Baseline model is

1164
00:49:54,000 --> 00:49:58,400
probably going to get around 90%

1165
00:49:55,599 --> 00:50:00,559
accuracy We Know by like eyeballing

1166
00:49:58,400 --> 00:50:06,240
eyeballing the data or something like

1167
00:50:00,559 --> 00:50:09,599
that and then we think our um we think

1168
00:50:06,240 --> 00:50:13,799
our model is going to get 93%

1169
00:50:09,599 --> 00:50:17,160
accuracy uh and we want a significant

1170
00:50:13,799 --> 00:50:19,440
threshold significance threshold of p is

1171
00:50:17,160 --> 00:50:22,319
less than

1172
00:50:19,440 --> 00:50:26,000
0.05 given these

1173
00:50:22,319 --> 00:50:30,559
two quantities we can basically go in

1174
00:50:26,000 --> 00:50:33,720
and say okay now we need uh 500 training

1175
00:50:30,559 --> 00:50:36,200
500 test examples in order to say with

1176
00:50:33,720 --> 00:50:38,920
confidence that we will be able

1177
00:50:36,200 --> 00:50:40,599
to um that we will be able to

1178
00:50:38,920 --> 00:50:42,640
distinguish between two models with 90

1179
00:50:40,599 --> 00:50:44,400
and 93%

1180
00:50:42,640 --> 00:50:48,240
accuracy

1181
00:50:44,400 --> 00:50:51,079
and I can go I can show the algorithm

1182
00:50:48,240 --> 00:50:51,079
that they have in this

1183
00:50:54,440 --> 00:50:57,440
paper

1184
00:51:01,760 --> 00:51:04,960
but basically the way this

1185
00:51:13,040 --> 00:51:19,720
works um is you sample a data set um

1186
00:51:17,799 --> 00:51:22,960
Canute the effect of interest on the

1187
00:51:19,720 --> 00:51:25,880
sample I compute the P value and then

1188
00:51:22,960 --> 00:51:29,319
you can calculate the power uh

1189
00:51:25,880 --> 00:51:31,520
by basically um checking the number of

1190
00:51:29,319 --> 00:51:34,480
times that the P value is less than your

1191
00:51:31,520 --> 00:51:36,319
threshold um multiplied by uh the fact

1192
00:51:34,480 --> 00:51:38,920
that the sign is in a particular

1193
00:51:36,319 --> 00:51:41,200
direction and by doing this you can

1194
00:51:38,920 --> 00:51:43,280
essentially um you can essentially

1195
00:51:41,200 --> 00:51:46,200
calculate how much data you would need

1196
00:51:43,280 --> 00:51:48,319
or sorry you can calculate the uh the

1197
00:51:46,200 --> 00:51:50,319
statistical power and then you can do

1198
00:51:48,319 --> 00:51:52,000
this for various sizes of data set so

1199
00:51:50,319 --> 00:51:53,559
you can gradually increase the size of

1200
00:51:52,000 --> 00:51:57,160
the data set or decrease the size of the

1201
00:51:53,559 --> 00:51:59,040
data set and that allows you to figure

1202
00:51:57,160 --> 00:52:02,200
out how big your data set needs to be in

1203
00:51:59,040 --> 00:52:04,640
order to get a statistically significant

1204
00:52:02,200 --> 00:52:08,839
effect of the data

1205
00:52:04,640 --> 00:52:10,720
set and so like many many people ask me

1206
00:52:08,839 --> 00:52:12,599
the question like how big of a data set

1207
00:52:10,720 --> 00:52:14,440
do we need to make this is basically the

1208
00:52:12,599 --> 00:52:17,280
statistically like quote unquote correct

1209
00:52:14,440 --> 00:52:19,520
answer for how you can do this and also

1210
00:52:17,280 --> 00:52:20,440
uh for assignment two we're going to ask

1211
00:52:19,520 --> 00:52:24,559
you to

1212
00:52:20,440 --> 00:52:26,720
justify uh your choice of creation of a

1213
00:52:24,559 --> 00:52:30,359
data set of particular size for testing

1214
00:52:26,720 --> 00:52:31,799
based on this so um uh pay pay attention

1215
00:52:30,359 --> 00:52:34,720
and please look at the references here

1216
00:52:31,799 --> 00:52:38,760
and you should be able to

1217
00:52:34,720 --> 00:52:41,280
that cool um any

1218
00:52:38,760 --> 00:52:43,119
questions I I didn't go like really

1219
00:52:41,280 --> 00:52:44,319
deeply into the formulas here you'll

1220
00:52:43,119 --> 00:52:45,720
you'll probably have to look them up in

1221
00:52:44,319 --> 00:52:48,119
the paper but hopefully that gives you

1222
00:52:45,720 --> 00:52:51,799
the general

1223
00:52:48,119 --> 00:52:52,680
idea okay next um how much training data

1224
00:52:51,799 --> 00:52:55,599
do I

1225
00:52:52,680 --> 00:52:58,160
need so in general more is usually

1226
00:52:55,599 --> 00:53:00,760
better if you're fine tuning a model um

1227
00:52:58,160 --> 00:53:02,880
so I can't tell you like you don't need

1228
00:53:00,760 --> 00:53:05,480
to make more data because

1229
00:53:02,880 --> 00:53:06,280
probably you do if you're not happy with

1230
00:53:05,480 --> 00:53:10,799
your

1231
00:53:06,280 --> 00:53:12,599
performance um but recently you can get

1232
00:53:10,799 --> 00:53:14,680
very reasonable performance with few

1233
00:53:12,599 --> 00:53:17,319
shot or zero shot or pre-trained models

1234
00:53:14,680 --> 00:53:19,760
and prompting and because of this in

1235
00:53:17,319 --> 00:53:21,240
some cases maybe the answer is zero

1236
00:53:19,760 --> 00:53:22,960
maybe you don't need any training data

1237
00:53:21,240 --> 00:53:26,559
and you could just use a zero shot pred

1238
00:53:22,960 --> 00:53:29,240
model so um you you need to choose like

1239
00:53:26,559 --> 00:53:31,319
what your accuracy threshold is um you

1240
00:53:29,240 --> 00:53:32,720
need to decide whether you want to be

1241
00:53:31,319 --> 00:53:34,480
fine-tuning a model to improve

1242
00:53:32,720 --> 00:53:36,319
performance or doing other things like

1243
00:53:34,480 --> 00:53:39,119
prompt engineering or other stuff like

1244
00:53:36,319 --> 00:53:41,520
that so basically there's no uh correct

1245
00:53:39,119 --> 00:53:45,440
answer to this

1246
00:53:41,520 --> 00:53:47,359
um one thing to be aware of is uh

1247
00:53:45,440 --> 00:53:51,440
sometimes if you select data

1248
00:53:47,359 --> 00:53:52,880
intelligently you can uh improve more

1249
00:53:51,440 --> 00:53:54,359
quickly with something like Active

1250
00:53:52,880 --> 00:53:56,520
Learning and active learning chooses

1251
00:53:54,359 --> 00:54:00,000
representative and difficult data that

1252
00:53:56,520 --> 00:54:02,559
you can um be

1253
00:54:00,000 --> 00:54:04,839
using so when you sample data for fine

1254
00:54:02,559 --> 00:54:07,440
tuning uh what you want to be doing is

1255
00:54:04,839 --> 00:54:08,839
you want to be sampling data that has

1256
00:54:07,440 --> 00:54:10,040
good coverage of the domains that you

1257
00:54:08,839 --> 00:54:12,760
want to

1258
00:54:10,040 --> 00:54:15,079
cover um you also want to be covering

1259
00:54:12,760 --> 00:54:18,599
for example language uh languages or

1260
00:54:15,079 --> 00:54:23,200
language varieties or demographics of

1261
00:54:18,599 --> 00:54:25,520
users um and another thing is uh when

1262
00:54:23,200 --> 00:54:29,440
you're doing this it's often good idea

1263
00:54:25,520 --> 00:54:31,400
to document how you're creating data and

1264
00:54:29,440 --> 00:54:34,079
uh there's this paper data statements

1265
00:54:31,400 --> 00:54:35,520
for NLP by vendor and fredman uh which

1266
00:54:34,079 --> 00:54:37,440
suggests a bunch of different things

1267
00:54:35,520 --> 00:54:39,520
that you can use to document your data

1268
00:54:37,440 --> 00:54:41,520
collection and like why and how you

1269
00:54:39,520 --> 00:54:44,960
collected the data and this gives you

1270
00:54:41,520 --> 00:54:47,200
some pieces of information that uh could

1271
00:54:44,960 --> 00:54:49,359
be useful this has been incorporated

1272
00:54:47,200 --> 00:54:51,880
into the hugging face data sets data set

1273
00:54:49,359 --> 00:54:53,520
cards and now hugging face data sets

1274
00:54:51,880 --> 00:54:56,040
actually has lots of metadata that's

1275
00:54:53,520 --> 00:54:58,359
kind of inspired by uh this although

1276
00:54:56,040 --> 00:55:01,799
it's been adjusted for more kind of like

1277
00:54:58,359 --> 00:55:01,799
practical industry use

1278
00:55:02,119 --> 00:55:06,480
cases another thing is annotation

1279
00:55:04,400 --> 00:55:09,160
guidelines so if you're asking humans to

1280
00:55:06,480 --> 00:55:11,319
do anything um or for that matter if

1281
00:55:09,160 --> 00:55:16,119
you're asking gp4 to generate data for

1282
00:55:11,319 --> 00:55:21,480
you um you need to tell people or gp4 in

1283
00:55:16,119 --> 00:55:24,440
um you know a clear manner how you will

1284
00:55:21,480 --> 00:55:28,119
um like how it should be creating data

1285
00:55:24,440 --> 00:55:29,920
so the first thing is um if you try uh

1286
00:55:28,119 --> 00:55:32,960
to an the first thing that you can do is

1287
00:55:29,920 --> 00:55:34,240
you can try to annotate yourself um and

1288
00:55:32,960 --> 00:55:37,039
if you actually try to solve The

1289
00:55:34,240 --> 00:55:38,440
annotation task yourself then you'll

1290
00:55:37,039 --> 00:55:41,160
realize that there's lots of corner

1291
00:55:38,440 --> 00:55:43,799
cases that are hard to decide on um

1292
00:55:41,160 --> 00:55:45,440
other things like that so like if you're

1293
00:55:43,799 --> 00:55:47,520
annotating sentiment what is the

1294
00:55:45,440 --> 00:55:49,799
boundary between very positive and

1295
00:55:47,520 --> 00:55:50,880
positive um if you're annotating

1296
00:55:49,799 --> 00:55:54,000
question

1297
00:55:50,880 --> 00:55:56,280
answering um like for

1298
00:55:54,000 --> 00:55:57,720
example do you want to answer in a whole

1299
00:55:56,280 --> 00:56:01,119
sentence or do you want to answer with

1300
00:55:57,720 --> 00:56:03,760
only a short concise answer like these

1301
00:56:01,119 --> 00:56:05,400
sorts of things you'll need to tell uh

1302
00:56:03,760 --> 00:56:07,839
either an annotator or a model that

1303
00:56:05,400 --> 00:56:10,960
you're asking to do annotation to give

1304
00:56:07,839 --> 00:56:12,760
some examples from pent Tree Bank uh

1305
00:56:10,960 --> 00:56:15,440
part of speech annotation guidelines

1306
00:56:12,760 --> 00:56:18,079
this is very old it's from 1990 but

1307
00:56:15,440 --> 00:56:21,200
basically they have uh like adverb this

1308
00:56:18,079 --> 00:56:25,559
category includes most words that end in

1309
00:56:21,200 --> 00:56:30,680
um ly as well as degree words like

1310
00:56:25,559 --> 00:56:33,079
quite um etc etc it has other things for

1311
00:56:30,680 --> 00:56:36,200
adverbs and then it has like confusing

1312
00:56:33,079 --> 00:56:38,039
parts of speech with examples uh one

1313
00:56:36,200 --> 00:56:39,640
thing that I found like really really

1314
00:56:38,039 --> 00:56:42,640
interesting is like if you look at these

1315
00:56:39,640 --> 00:56:46,160
annotation guidelines it's like uh

1316
00:56:42,640 --> 00:56:48,319
prompts so if you look at this it's like

1317
00:56:46,160 --> 00:56:49,880
these are your your prompts your zero

1318
00:56:48,319 --> 00:56:52,359
shot prompts and these are F shot

1319
00:56:49,880 --> 00:56:54,480
examples so like even for humans we were

1320
00:56:52,359 --> 00:56:56,520
doing F shot prompting with examples

1321
00:56:54,480 --> 00:57:00,880
when they were doing annotations so uh

1322
00:56:56,520 --> 00:57:03,119
it's kind of uh kind of fun um hiring

1323
00:57:00,880 --> 00:57:05,000
annotators so like let's say you want to

1324
00:57:03,119 --> 00:57:08,319
actually build a data set and and pay

1325
00:57:05,000 --> 00:57:10,359
people to do things um for smaller scale

1326
00:57:08,319 --> 00:57:13,359
projects uh very often you can just

1327
00:57:10,359 --> 00:57:15,240
annotate yourself and that's fine um

1328
00:57:13,359 --> 00:57:16,720
there's a fixed set of overhead to get

1329
00:57:15,240 --> 00:57:19,480
other people to do something and train

1330
00:57:16,720 --> 00:57:23,200
them and stuff so you know I often just

1331
00:57:19,480 --> 00:57:25,079
annotate things myself um you can also

1332
00:57:23,200 --> 00:57:26,520
find friends or other students or

1333
00:57:25,079 --> 00:57:29,559
co-workers who can help you out with

1334
00:57:26,520 --> 00:57:33,359
things you can bri bribe them with uh

1335
00:57:29,559 --> 00:57:37,280
pizza or whatever favorite uh food or

1336
00:57:33,359 --> 00:57:39,400
beverage that they like um then for

1337
00:57:37,280 --> 00:57:42,440
finding people online there's a lot of

1338
00:57:39,400 --> 00:57:45,160
things that you can do um I very often

1339
00:57:42,440 --> 00:57:46,000
hire Freelancers uh through platforms

1340
00:57:45,160 --> 00:57:50,400
such as

1341
00:57:46,000 --> 00:57:51,799
upwork um this is good and bad the bad

1342
00:57:50,400 --> 00:57:53,760
thing about it is that this is often

1343
00:57:51,799 --> 00:57:56,280
more expensive the good thing about it

1344
00:57:53,760 --> 00:57:58,640
is um you get people who have pride in

1345
00:57:56,280 --> 00:58:00,440
their work and accountability and

1346
00:57:58,640 --> 00:58:02,440
motivation because like if they get

1347
00:58:00,440 --> 00:58:04,480
rated poorly they it's going to be

1348
00:58:02,440 --> 00:58:06,720
harder to get work and often they're

1349
00:58:04,480 --> 00:58:08,160
Professionals in their fields so like if

1350
00:58:06,720 --> 00:58:12,079
you want to get a code generation data

1351
00:58:08,160 --> 00:58:15,880
set you can hire good um Freelancers

1352
00:58:12,079 --> 00:58:18,520
I've actually heard rumors that uh

1353
00:58:15,880 --> 00:58:20,119
people like open AI they hire people and

1354
00:58:18,520 --> 00:58:21,599
pay them $60 an hour to do The

1355
00:58:20,119 --> 00:58:23,599
annotation because they really want

1356
00:58:21,599 --> 00:58:27,119
people who are very professional and do

1357
00:58:23,599 --> 00:58:30,000
a very good job um I don't pay that

1358
00:58:27,119 --> 00:58:34,240
much but I do pay well more than minimum

1359
00:58:30,000 --> 00:58:35,880
wage and uh you know like it's a I pay a

1360
00:58:34,240 --> 00:58:38,039
competitive price for these freelancing

1361
00:58:35,880 --> 00:58:40,319
sites when I get people to do

1362
00:58:38,039 --> 00:58:42,000
that another thing you can do as crowd

1363
00:58:40,319 --> 00:58:44,400
workers and this is could be through

1364
00:58:42,000 --> 00:58:45,960
sites like Mechanical Turk or prolific

1365
00:58:44,400 --> 00:58:48,960
or other things like this so that's

1366
00:58:45,960 --> 00:58:51,680
another option um here quality control

1367
00:58:48,960 --> 00:58:55,240
becomes very difficult and um we're

1368
00:58:51,680 --> 00:58:57,799
getting to the point where number one

1369
00:58:55,240 --> 00:58:59,400
um if you don't aren't very careful with

1370
00:58:57,799 --> 00:59:01,920
quality control language models actually

1371
00:58:59,400 --> 00:59:03,400
do a similarly good job as crowd workers

1372
00:59:01,920 --> 00:59:06,960
and number two all the crowd workers are

1373
00:59:03,400 --> 00:59:10,000
using gp4 anyway so um you do need to be

1374
00:59:06,960 --> 00:59:12,319
careful about that um one thing that I

1375
00:59:10,000 --> 00:59:14,039
often do is I hire for a small job first

1376
00:59:12,319 --> 00:59:16,880
to gauge timeliness and accuracy and

1377
00:59:14,039 --> 00:59:18,920
then hire for a bigger job so um just

1378
00:59:16,880 --> 00:59:21,720
hire people to do you know 50 examples

1379
00:59:18,920 --> 00:59:23,319
or 20 examples first and then uh you

1380
00:59:21,720 --> 00:59:26,240
know if they do a good job with it then

1381
00:59:23,319 --> 00:59:27,960
I hire them to do 200 th000

1382
00:59:26,240 --> 00:59:30,799
examples

1383
00:59:27,960 --> 00:59:34,720
um one thing to note is that if you're

1384
00:59:30,799 --> 00:59:36,599
doing research in a university um you

1385
00:59:34,720 --> 00:59:39,400
might need to get approval from an

1386
00:59:36,599 --> 00:59:41,480
Institutional review board and this is

1387
00:59:39,400 --> 00:59:43,000
in particular the case for subjective

1388
00:59:41,480 --> 00:59:45,880
task so this is when you're asking

1389
00:59:43,000 --> 00:59:47,440
people how do you feel about this output

1390
00:59:45,880 --> 00:59:50,039
um do you think this output is

1391
00:59:47,440 --> 00:59:51,720
representative of your beliefs or things

1392
00:59:50,039 --> 00:59:54,760
like that where it doesn't have a

1393
00:59:51,720 --> 00:59:56,319
correct answer a yes and no answer if

1394
00:59:54,760 --> 00:59:58,680
it's something like it it does have a

1395
00:59:56,319 --> 01:00:03,640
yes and no answer which is like how many

1396
00:59:58,680 --> 01:00:05,640
verbs are in this sentence or um how do

1397
01:00:03,640 --> 01:00:07,280
you translate the sentence into another

1398
01:00:05,640 --> 01:00:09,880
language or something like that then you

1399
01:00:07,280 --> 01:00:12,039
don't need an IRB approval um but if

1400
01:00:09,880 --> 01:00:15,000
it's borderline you might want to check

1401
01:00:12,039 --> 01:00:17,280
anyway um so that that's something to be

1402
01:00:15,000 --> 01:00:17,280
aware

1403
01:00:18,640 --> 01:00:26,240
of next is assessing annotation quality

1404
01:00:22,640 --> 01:00:27,680
so um one of my favorite ways to do this

1405
01:00:26,240 --> 01:00:30,039
is assess Human

1406
01:00:27,680 --> 01:00:32,240
Performance and so the way we do this is

1407
01:00:30,039 --> 01:00:34,119
you double annotate some data and then

1408
01:00:32,240 --> 01:00:37,160
you measure whatever metric you want to

1409
01:00:34,119 --> 01:00:39,200
measure for machines just with respect

1410
01:00:37,160 --> 01:00:41,039
to human agreement and so for

1411
01:00:39,200 --> 01:00:43,839
translation if you're using blue score

1412
01:00:41,039 --> 01:00:45,440
or KF score or something like this then

1413
01:00:43,839 --> 01:00:47,079
you would want to use this for

1414
01:00:45,440 --> 01:00:50,440
assessment of the

1415
01:00:47,079 --> 01:00:56,039
outputs um the advantage of doing this

1416
01:00:50,440 --> 01:00:58,760
is that you get a human quality score

1417
01:00:56,039 --> 01:01:00,960
and the human quality score is directly

1418
01:00:58,760 --> 01:01:02,480
comparable to the machine quality score

1419
01:01:00,960 --> 01:01:04,599
and so you can say well humans got the

1420
01:01:02,480 --> 01:01:07,280
task right 90% of the time and gp4 got

1421
01:01:04,599 --> 01:01:11,280
the task right 16% of the time so humans

1422
01:01:07,280 --> 01:01:13,760
are way better than gp4 or um you know

1423
01:01:11,280 --> 01:01:16,559
humans got it right 80% of the time and

1424
01:01:13,760 --> 01:01:19,599
gp4 got it right 78% of the time so this

1425
01:01:16,559 --> 01:01:21,000
task is you know this task or maybe not

1426
01:01:19,599 --> 01:01:23,640
necessarily the task but at least the

1427
01:01:21,000 --> 01:01:25,079
data set is more or less uh been so by

1428
01:01:23,640 --> 01:01:26,640
the strongest language models so now we

1429
01:01:25,079 --> 01:01:28,920
need to catch up open source models so

1430
01:01:26,640 --> 01:01:31,680
SW ones or something like

1431
01:01:28,920 --> 01:01:32,880
that um there are things that you can

1432
01:01:31,680 --> 01:01:34,880
measure you can measure things like

1433
01:01:32,880 --> 01:01:36,880
Kappa statistics this is particularly

1434
01:01:34,880 --> 01:01:39,799
useful for um kind of just

1435
01:01:36,880 --> 01:01:41,799
classification tasks and what this tells

1436
01:01:39,799 --> 01:01:43,880
you is this tells you how much higher is

1437
01:01:41,799 --> 01:01:48,000
the agreement that you would get than if

1438
01:01:43,880 --> 01:01:49,920
you got it by chance and so for example

1439
01:01:48,000 --> 01:01:53,279
let's say you're classifying

1440
01:01:49,920 --> 01:01:54,760
spam uh or you're classifying you know

1441
01:01:53,279 --> 01:01:59,520
toxic content or something something

1442
01:01:54,760 --> 01:02:03,400
like that in 99% of your time 99% of the

1443
01:01:59,520 --> 01:02:07,480
time the content is not toxic and 1% of

1444
01:02:03,400 --> 01:02:11,799
the time the content is toxic and then

1445
01:02:07,480 --> 01:02:14,079
you hire some annotators and you get 98%

1446
01:02:11,799 --> 01:02:16,279
accuracy that's kind of bad right you

1447
01:02:14,079 --> 01:02:19,200
know if you just said not toxic all the

1448
01:02:16,279 --> 01:02:20,880
time you would get 99% um what the Kaus

1449
01:02:19,200 --> 01:02:24,599
statistic does is it accounts for this

1450
01:02:20,880 --> 01:02:26,559
basically it says um how much more like

1451
01:02:24,599 --> 01:02:28,440
assis than chance and if you just had

1452
01:02:26,559 --> 01:02:30,720
chance accuracy you would get zero if

1453
01:02:28,440 --> 01:02:33,200
you had perfect accuracy you would get

1454
01:02:30,720 --> 01:02:34,920
one and you normally get something in

1455
01:02:33,200 --> 01:02:37,359
between

1456
01:02:34,920 --> 01:02:39,200
um so if it's slow you may need to

1457
01:02:37,359 --> 01:02:41,319
revisit guidelines Tire better

1458
01:02:39,200 --> 01:02:44,480
annotators or rethink whether the task

1459
01:02:41,319 --> 01:02:46,559
is possible at all or not um and you

1460
01:02:44,480 --> 01:02:48,599
know some tasks are just impossible like

1461
01:02:46,559 --> 01:02:51,599
if um I'm

1462
01:02:48,599 --> 01:02:51,599
asking

1463
01:02:52,240 --> 01:02:58,160
uh well or um they're very hard for

1464
01:02:55,960 --> 01:03:00,039
annotators so like to give one example

1465
01:02:58,160 --> 01:03:04,039
um annotators are really horrible at

1466
01:03:00,039 --> 01:03:06,200
identifying fake reviews um and so like

1467
01:03:04,039 --> 01:03:07,640
if you even if you hire annotators to

1468
01:03:06,200 --> 01:03:09,279
identify paper reviews they're bad at

1469
01:03:07,640 --> 01:03:11,359
doing that so you're not likely to get

1470
01:03:09,279 --> 01:03:14,680
high

1471
01:03:11,359 --> 01:03:17,920
agreement um cool I'm going to skip over

1472
01:03:14,680 --> 01:03:23,279
this part because I already talked about

1473
01:03:17,920 --> 01:03:26,640
it okay um any any questions

1474
01:03:23,279 --> 01:03:29,079
here okay sounds good uh next I'd like

1475
01:03:26,640 --> 01:03:30,640
to get into running experiments so

1476
01:03:29,079 --> 01:03:34,359
running experiments one thing I find

1477
01:03:30,640 --> 01:03:37,200
very helpful is workflow automation um

1478
01:03:34,359 --> 01:03:40,079
and basically what I I like to do is I

1479
01:03:37,200 --> 01:03:41,839
like to mod modularize each step of an

1480
01:03:40,079 --> 01:03:44,119
experiment into a

1481
01:03:41,839 --> 01:03:47,240
directory

1482
01:03:44,119 --> 01:03:51,039
um where uh you have like a directory as

1483
01:03:47,240 --> 01:03:53,279
input and a directory as output

1484
01:03:51,039 --> 01:03:54,559
um this is my personal way of doing

1485
01:03:53,279 --> 01:03:56,799
things there are other ways of doing

1486
01:03:54,559 --> 01:03:58,640
things that are also good but um very

1487
01:03:56,799 --> 01:04:00,760
often like just to give an example

1488
01:03:58,640 --> 01:04:04,680
you'll need to do pre-processing

1489
01:04:00,760 --> 01:04:07,480
According to some uh you'll need to do

1490
01:04:04,680 --> 01:04:09,119
data selection so you'll need to select

1491
01:04:07,480 --> 01:04:11,039
which data sets you're training on

1492
01:04:09,119 --> 01:04:13,520
you'll need to do pre-processing of them

1493
01:04:11,039 --> 01:04:16,160
with a tokenization model and then you

1494
01:04:13,520 --> 01:04:18,359
will need to run an

1495
01:04:16,160 --> 01:04:20,000
experiment and then you'll need to do

1496
01:04:18,359 --> 01:04:23,240
evaluation and those are all kind of

1497
01:04:20,000 --> 01:04:25,079
like discret Steps where the data

1498
01:04:23,240 --> 01:04:27,760
selection takes in your big pool of data

1499
01:04:25,079 --> 01:04:31,200
and outputs a a data set that's been

1500
01:04:27,760 --> 01:04:33,680
selected the tokenization

1501
01:04:31,200 --> 01:04:35,480
will uh take a tokenizer model maybe

1502
01:04:33,680 --> 01:04:38,599
train a tokenizer model and and split it

1503
01:04:35,480 --> 01:04:40,400
up into different tokens um the training

1504
01:04:38,599 --> 01:04:42,079
will train it might output a whole bunch

1505
01:04:40,400 --> 01:04:44,720
of checkpoints and the evaluation will

1506
01:04:42,079 --> 01:04:47,039
evaluate one checkpoint and so those are

1507
01:04:44,720 --> 01:04:48,400
all kind of modular and you can actually

1508
01:04:47,039 --> 01:04:50,039
think of each one of them as like a

1509
01:04:48,400 --> 01:04:52,760
function in your Python

1510
01:04:50,039 --> 01:04:56,400
program

1511
01:04:52,760 --> 01:04:58,160
and you kind of want to avoid rerunning

1512
01:04:56,400 --> 01:05:00,200
data set selection and tokenization

1513
01:04:58,160 --> 01:05:01,720
every time you do a new evaluation right

1514
01:05:00,200 --> 01:05:03,359
like that would be kind of silly you

1515
01:05:01,720 --> 01:05:04,680
definitely want to avoid rerunning

1516
01:05:03,359 --> 01:05:09,119
training every time you evaluate a

1517
01:05:04,680 --> 01:05:11,200
checkpoint so um what I do is I often

1518
01:05:09,119 --> 01:05:12,799
name directories by parameters where

1519
01:05:11,200 --> 01:05:16,079
it's like Transformer

1520
01:05:12,799 --> 01:05:18,640
layer Transformer layer 8 node 512

1521
01:05:16,079 --> 01:05:21,279
Dropout 0.5 label smooth

1522
01:05:18,640 --> 01:05:25,880
0.02 um and so I have all the parameters

1523
01:05:21,279 --> 01:05:26,880
in there and then

1524
01:05:25,880 --> 01:05:29,680
the

1525
01:05:26,880 --> 01:05:31,960
training process will output a whole

1526
01:05:29,680 --> 01:05:33,960
bunch of checkpoints in here and then

1527
01:05:31,960 --> 01:05:35,520
for my evaluation I have evaluation

1528
01:05:33,960 --> 01:05:38,119
metrics and I have the checkpoint I'm

1529
01:05:35,520 --> 01:05:41,680
evaluating so uh when I do

1530
01:05:38,119 --> 01:05:45,119
evaluation I will then append checkpoint

1531
01:05:41,680 --> 01:05:47,279
6 uh metric F measure or something like

1532
01:05:45,119 --> 01:05:49,079
that and so I keep around all of the

1533
01:05:47,279 --> 01:05:52,520
previous information and just append

1534
01:05:49,079 --> 01:05:54,599
append append append and so um this

1535
01:05:52,520 --> 01:05:56,680
allows you to avoid rerunning things

1536
01:05:54,599 --> 01:05:58,359
because you can uh just have your python

1537
01:05:56,680 --> 01:06:00,520
code to check if the directory already

1538
01:05:58,359 --> 01:06:01,839
exists and already has been completed

1539
01:06:00,520 --> 01:06:03,559
and then read in the result if it

1540
01:06:01,839 --> 01:06:06,319
already has been or run the experiment

1541
01:06:03,559 --> 01:06:08,079
that it hasn't been so um you can write

1542
01:06:06,319 --> 01:06:10,279
you can write this in pure python by

1543
01:06:08,079 --> 01:06:11,599
just adding like some if statements at

1544
01:06:10,279 --> 01:06:14,079
the beginning of the function some if

1545
01:06:11,599 --> 01:06:16,799
statements at um some like output

1546
01:06:14,079 --> 01:06:19,440
statements at the end of the function um

1547
01:06:16,799 --> 01:06:22,000
there are more sophisticated models

1548
01:06:19,440 --> 01:06:24,200
methods so there's like a toolkit called

1549
01:06:22,000 --> 01:06:28,079
duct tape that was originally created

1550
01:06:24,200 --> 01:06:31,760
here at CMU and um my uh student Patrick

1551
01:06:28,079 --> 01:06:33,079
is maintaining now this link um so you

1552
01:06:31,760 --> 01:06:34,960
can either just roll something on your

1553
01:06:33,079 --> 01:06:36,880
own or look into one of these more

1554
01:06:34,960 --> 01:06:39,359
complex work workflow automation things

1555
01:06:36,880 --> 01:06:39,359
to sve you

1556
01:06:39,400 --> 01:06:47,279
time okay evaluation um so I talked

1557
01:06:43,400 --> 01:06:49,000
about this to some extent um uh so yeah

1558
01:06:47,279 --> 01:06:51,000
I'll just skip over

1559
01:06:49,000 --> 01:06:54,559
that

1560
01:06:51,000 --> 01:06:57,200
and result reporting um

1561
01:06:54,559 --> 01:06:59,160
for papers one thing that I really like

1562
01:06:57,200 --> 01:07:01,960
to do is plan the result section in

1563
01:06:59,160 --> 01:07:07,039
advance or at least imagine the result

1564
01:07:01,960 --> 01:07:07,039
section in advance um

1565
01:07:07,200 --> 01:07:11,640
so what what I think of is like what

1566
01:07:09,559 --> 01:07:14,520
experimental claims would I like to make

1567
01:07:11,640 --> 01:07:15,760
how am I going to support them by the

1568
01:07:14,520 --> 01:07:19,039
experiments that I'm going to show in a

1569
01:07:15,760 --> 01:07:21,160
result section um and this identifies

1570
01:07:19,039 --> 01:07:24,640
unjustified experimental claims like so

1571
01:07:21,160 --> 01:07:27,119
let's say your method is you're saying

1572
01:07:24,640 --> 01:07:29,000
something like uh this method improves

1573
01:07:27,119 --> 01:07:30,440
across a wide variety of languages and

1574
01:07:29,000 --> 01:07:32,520
then you realize that you only have one

1575
01:07:30,440 --> 01:07:34,720
language and you're uh in your

1576
01:07:32,520 --> 01:07:37,960
experiment section that's a problem

1577
01:07:34,720 --> 01:07:40,640
obviously um also I I really enjoy like

1578
01:07:37,960 --> 01:07:43,599
assuming that all of my experiments are

1579
01:07:40,640 --> 01:07:46,520
going really really well um and you know

1580
01:07:43,599 --> 01:07:49,440
none of my uh none of my runs crash with

1581
01:07:46,520 --> 01:07:52,000
Cuda out of memory errors and you know

1582
01:07:49,440 --> 01:07:55,319
all of all of the experiments appear as

1583
01:07:52,000 --> 01:07:57,960
expected and if you do something like

1584
01:07:55,319 --> 01:07:59,960
that you can be ambitious and say okay

1585
01:07:57,960 --> 01:08:03,119
how can I make this research project

1586
01:07:59,960 --> 01:08:04,960
really impactful like um and another

1587
01:08:03,119 --> 01:08:08,240
thing that I like to ask my students or

1588
01:08:04,960 --> 01:08:11,200
people I'm working with recently is like

1589
01:08:08,240 --> 01:08:13,440
who are like three people in the world

1590
01:08:11,200 --> 01:08:17,440
who will be really excited by your paper

1591
01:08:13,440 --> 01:08:19,040
like name actual people um and where do

1592
01:08:17,440 --> 01:08:20,839
those people work what do they care

1593
01:08:19,040 --> 01:08:22,359
about what sort of evidence would you

1594
01:08:20,839 --> 01:08:24,560
need in your paper to make them really

1595
01:08:22,359 --> 01:08:26,560
excited about your paper or something

1596
01:08:24,560 --> 01:08:29,679
like that and very often people will

1597
01:08:26,560 --> 01:08:31,480
reply to me like oh I think people in um

1598
01:08:29,679 --> 01:08:32,799
in Google will be very excited about

1599
01:08:31,480 --> 01:08:34,440
this and they're going to use it and I'm

1600
01:08:32,799 --> 01:08:38,719
like well you're writing all your code

1601
01:08:34,440 --> 01:08:39,839
in pytorch and they don't use pytorch so

1602
01:08:38,719 --> 01:08:41,000
how are you going to convince them to

1603
01:08:39,839 --> 01:08:42,640
use their paper they're going to have to

1604
01:08:41,000 --> 01:08:46,120
reimplement it in Jax and that's going

1605
01:08:42,640 --> 01:08:47,520
to suck for them so like uh you know

1606
01:08:46,120 --> 01:08:49,040
what are the barriers for them actually

1607
01:08:47,520 --> 01:08:50,799
using it and then maybe the people are

1608
01:08:49,040 --> 01:08:52,159
like oh well maybe actually I don't want

1609
01:08:50,799 --> 01:08:54,199
people at Google to use this and I can

1610
01:08:52,159 --> 01:08:56,560
think of somebody else and it's like

1611
01:08:54,199 --> 01:08:58,920
well great so now release it open source

1612
01:08:56,560 --> 01:09:00,520
and people will will have it open source

1613
01:08:58,920 --> 01:09:01,920
so you can kind of think about like the

1614
01:09:00,520 --> 01:09:03,719
types of evidence that you would need to

1615
01:09:01,920 --> 01:09:05,440
convince people to use your work and

1616
01:09:03,719 --> 01:09:08,040
that can result in your work being more

1617
01:09:05,440 --> 01:09:09,319
impactful in the long run and if you

1618
01:09:08,040 --> 01:09:10,400
think about it from the very beginning

1619
01:09:09,319 --> 01:09:11,839
that also helps you plan your

1620
01:09:10,400 --> 01:09:13,520
experiments like what sort of evidence

1621
01:09:11,839 --> 01:09:15,359
is necessary for people to get excited

1622
01:09:13,520 --> 01:09:18,440
about it in the this

1623
01:09:15,359 --> 01:09:20,120
SPS um another thing that I like to do

1624
01:09:18,440 --> 01:09:24,000
with result reporting is result

1625
01:09:20,120 --> 01:09:26,880
generation scripts um so uh I often

1626
01:09:24,000 --> 01:09:29,159
generate paper latex directly from log

1627
01:09:26,880 --> 01:09:31,799
files uh there's two reasons why I do

1628
01:09:29,159 --> 01:09:34,480
this um number one it's efficient and

1629
01:09:31,799 --> 01:09:36,719
minimizes errors number two it allows

1630
01:09:34,480 --> 01:09:39,080
you to preemptively plan experiments

1631
01:09:36,719 --> 01:09:41,120
that you want to run so like for example

1632
01:09:39,080 --> 01:09:44,440
if we go back to the dock um the

1633
01:09:41,120 --> 01:09:46,199
directory that I talked about before um

1634
01:09:44,440 --> 01:09:50,359
I can write

1635
01:09:46,199 --> 01:09:52,719
a a script that reads in 20 evaluation

1636
01:09:50,359 --> 01:09:54,800
results from 20 different directories

1637
01:09:52,719 --> 01:09:56,920
and fills in a table and if that

1638
01:09:54,800 --> 01:09:58,600
directory doesn't exist yet it will put

1639
01:09:56,920 --> 01:10:01,239
like TVD or something like that in the

1640
01:09:58,600 --> 01:10:03,960
table so I can very quickly see okay

1641
01:10:01,239 --> 01:10:05,880
these things are TBD um oh this thing

1642
01:10:03,960 --> 01:10:07,480
has been TBD for a very long time is my

1643
01:10:05,880 --> 01:10:09,400
experiment crashed do I need to go back

1644
01:10:07,480 --> 01:10:12,239
and like restart my experiment or

1645
01:10:09,400 --> 01:10:13,719
something like that so um it's an

1646
01:10:12,239 --> 01:10:17,280
efficient way and when you finish the

1647
01:10:13,719 --> 01:10:17,280
last TBD it's a very good feeling

1648
01:10:18,280 --> 01:10:23,719
also cool um next computational

1649
01:10:21,760 --> 01:10:26,159
resources actually I kind of already

1650
01:10:23,719 --> 01:10:28,600
talked about this a little bit um but on

1651
01:10:26,159 --> 01:10:30,280
Amazon web services we have uh class

1652
01:10:28,600 --> 01:10:32,080
credits that we're going to be issuing

1653
01:10:30,280 --> 01:10:34,880
as soon as uh the assignment one

1654
01:10:32,080 --> 01:10:37,560
deadline is over um there's also Google

1655
01:10:34,880 --> 01:10:39,440
cloud and collab um you can get

1656
01:10:37,560 --> 01:10:44,000
commodity gpus and other things like

1657
01:10:39,440 --> 01:10:47,800
that so um you can also consider

1658
01:10:44,000 --> 01:10:53,159
that okay let me get into Data analysis

1659
01:10:47,800 --> 01:10:55,440
um so I'm going to cover this a lot more

1660
01:10:53,159 --> 01:10:58,480
in an interpretation lecture and this is

1661
01:10:55,440 --> 01:10:59,520
going to be in three classes so this is

1662
01:10:58,480 --> 01:11:02,239
going to

1663
01:10:59,520 --> 01:11:07,000
be the

1664
01:11:02,239 --> 01:11:09,719
Tuesday after next um so uh very

1665
01:11:07,000 --> 01:11:11,000
important things though uh look at data

1666
01:11:09,719 --> 01:11:13,679
um you'll want to do quantitative

1667
01:11:11,000 --> 01:11:16,239
analysis and qualitative analysis um you

1668
01:11:13,679 --> 01:11:17,440
can also look at model explanations so

1669
01:11:16,239 --> 01:11:18,719
I'm going to cover how to do all of

1670
01:11:17,440 --> 01:11:21,520
these things in that lecture I don't

1671
01:11:18,719 --> 01:11:24,440
have enough time to do it

1672
01:11:21,520 --> 01:11:26,960
today then the final thing is accoring

1673
01:11:24,440 --> 01:11:30,840
conclusions um this is also too much for

1674
01:11:26,960 --> 01:11:34,000
a single class but um I very highly

1675
01:11:30,840 --> 01:11:35,920
recommend this lecture um uh sorry these

1676
01:11:34,000 --> 01:11:39,320
lecture slides they don't take that long

1677
01:11:35,920 --> 01:11:40,880
to look through they're maybe um 20

1678
01:11:39,320 --> 01:11:42,880
minutes or so but they're very very

1679
01:11:40,880 --> 01:11:45,480
helpful um they talk about how to

1680
01:11:42,880 --> 01:11:48,199
structure a paper uh other things like

1681
01:11:45,480 --> 01:11:51,440
this and if you follow this advice for

1682
01:11:48,199 --> 01:11:53,239
writing your reports for like three and

1683
01:11:51,440 --> 01:11:54,960
four assignment three and assignment

1684
01:11:53,239 --> 01:11:57,800
four even assignment two I think you

1685
01:11:54,960 --> 01:11:59,400
can't really go wrong uh actually three

1686
01:11:57,800 --> 01:12:00,840
and four is probably better uh than

1687
01:11:59,400 --> 01:12:03,320
assignment two assignment two can be

1688
01:12:00,840 --> 01:12:05,360
more descriptive so definitely take a

1689
01:12:03,320 --> 01:12:08,600
look at that if

1690
01:12:05,360 --> 01:12:08,600
you cool