1
00:00:00,080 --> 00:00:02,719
Speaker 1: In this episode, we'll be focusing on Project Debater, which

2
00:00:02,720 --> 00:00:06,800
Speaker 1: is an AI system designed to process evidence and persuasive

3
00:00:06,880 --> 00:00:10,039
Speaker 1: arguments and text so that it can ultimately understand and

4
00:00:10,080 --> 00:00:13,800
Speaker 1: participate in human debate. To get to the heart of

5
00:00:13,840 --> 00:00:16,880
Speaker 1: this effort, we're going to share two interviews we recorded

6
00:00:16,880 --> 00:00:20,799
Speaker 1: with leaders at IBM. The first is with Noam slow Name,

7
00:00:21,120 --> 00:00:24,200
Speaker 1: who is a distinguished engineer at IBM Research and founder

8
00:00:24,239 --> 00:00:27,600
Speaker 1: of Project Debater, and the second chat will be with

9
00:00:27,680 --> 00:00:31,840
Speaker 1: matdou Coachar, who is Vice President Offering Management for IBM

10
00:00:31,920 --> 00:00:34,680
Speaker 1: Data and AI. So today's episode is going to be

11
00:00:34,760 --> 00:00:38,000
Speaker 1: the third of four episodes in this series that Robert

12
00:00:38,040 --> 00:00:39,840
Speaker 1: and I are releasing here on the Stuff to Blow

13
00:00:39,840 --> 00:00:42,720
Speaker 1: Your Mind feed. If you'd like to hear more episodes,

14
00:00:42,800 --> 00:00:45,159
Speaker 1: you can check out the ones labeled smart Talks that

15
00:00:45,200 --> 00:00:47,479
Speaker 1: we've released over the past few weeks, and you can

16
00:00:47,520 --> 00:00:50,480
Speaker 1: also listen to the first four episodes of smart Talks,

17
00:00:50,520 --> 00:00:52,840
Speaker 1: which were released not on our show but in the feed.

18
00:00:52,880 --> 00:00:55,400
Speaker 1: For the podcast Text Stuff. You can find them on

19
00:00:55,440 --> 00:00:57,720
Speaker 1: the I Heart Radio app or wherever you get your podcast.

20
00:00:57,800 --> 00:01:00,000
Speaker 1: Just look up text Stuff and click on the episode

21
00:01:00,040 --> 00:01:02,840
Speaker 1: has labeled Smart Talks, and of course stay tuned for

22
00:01:02,840 --> 00:01:04,920
Speaker 1: the one remaining episode in the series, which is going

23
00:01:05,000 --> 00:01:06,839
Speaker 1: to be published in our feed in a couple of weeks.

24
00:01:07,200 --> 00:01:10,320
Speaker 1: And now straight onto our conversation with no One Slowly,

25
00:01:12,920 --> 00:01:14,880
Speaker 1: no One, thanks so much for joining us today. Can

26
00:01:14,920 --> 00:01:18,360
Speaker 1: you start by introducing yourself and talking about your role

27
00:01:18,400 --> 00:01:22,880
Speaker 1: at IBM? Sure, thank you for hosting me. So I'm

28
00:01:22,920 --> 00:01:26,200
Speaker 1: no One Slownym. I'm a distinguished engineer at IBMI Research.

29
00:01:27,440 --> 00:01:30,920
Speaker 1: I did my PhD in the Hebew University quite a

30
00:01:30,920 --> 00:01:36,400
Speaker 1: few years ago walking on machine learning, staff and artificial intelligence,

31
00:01:37,040 --> 00:01:39,600
Speaker 1: and then I did a past doc at Princeton University

32
00:01:39,680 --> 00:01:44,240
Speaker 1: and I joined the IBM research in two thousand and seven,

33
00:01:45,319 --> 00:01:50,960
Speaker 1: and uh in two thousand and eleven, I suggested the

34
00:01:50,960 --> 00:01:53,320
Speaker 1: project that I guess we're going to talk about today,

35
00:01:53,520 --> 00:01:56,440
Speaker 1: and of course that project was Project Debat, right do

36
00:01:56,480 --> 00:01:58,280
Speaker 1: you do? You want to mention a little bit about

37
00:01:58,280 --> 00:02:01,640
Speaker 1: the origins of that. In IBM research, we have this

38
00:02:02,120 --> 00:02:08,840
Speaker 1: interesting tradition of grand challenges in artificial intelligence. Back in

39
00:02:08,919 --> 00:02:12,440
Speaker 1: the nineties, idem introduced the Blue that was able to

40
00:02:12,480 --> 00:02:16,800
Speaker 1: defeat Gary customers in chess, and in two thousand eleven

41
00:02:16,880 --> 00:02:19,720
Speaker 1: id AM introduced Watson that was able to defeat the

42
00:02:19,760 --> 00:02:23,760
Speaker 1: all time winners of the TV trivia game Jeopardy. And

43
00:02:23,919 --> 00:02:27,440
Speaker 1: just a few days after this event, an email was

44
00:02:27,520 --> 00:02:31,120
Speaker 1: sent to all the thousands of researchers in i DM

45
00:02:31,200 --> 00:02:35,880
Speaker 1: across the globe, myself included, asking us what should be

46
00:02:35,919 --> 00:02:41,119
Speaker 1: the next grand challenge for IDM research and uh I

47
00:02:41,160 --> 00:02:44,160
Speaker 1: was intrigued by that, so I offered my office mate

48
00:02:44,600 --> 00:02:48,480
Speaker 1: at the time to brainstone together, and this is what

49
00:02:48,520 --> 00:02:50,680
Speaker 1: we did. We set in the office in Tel Aviv

50
00:02:50,880 --> 00:02:54,440
Speaker 1: and we raised many different ideas that probably I should

51
00:02:54,480 --> 00:02:58,600
Speaker 1: not share with you today, but at some point towards

52
00:02:58,639 --> 00:03:02,640
Speaker 1: the end of the hour, well I suggested this notion

53
00:03:03,160 --> 00:03:08,320
Speaker 1: of developing a machine that we'll be able to debate humans,

54
00:03:08,520 --> 00:03:11,399
Speaker 1: and that this is how we will demonstrate the technology

55
00:03:11,480 --> 00:03:15,440
Speaker 1: for a full life debate between this envisioned system and

56
00:03:15,520 --> 00:03:20,720
Speaker 1: an expert human debate. And we submitted that the only

57
00:03:20,720 --> 00:03:23,240
Speaker 1: guidance that we got from the management was really to

58
00:03:23,280 --> 00:03:27,200
Speaker 1: submit the proposals in a single side so they will

59
00:03:27,240 --> 00:03:30,320
Speaker 1: not be swamped with too many details. And we were

60
00:03:30,400 --> 00:03:34,400
Speaker 1: able to helpfully follow these guidelines and we submitted a

61
00:03:34,440 --> 00:03:37,120
Speaker 1: single slide. This was fair Boy in two thousand eleven,

62
00:03:37,840 --> 00:03:40,920
Speaker 1: and this started a fairly long, and the thought review

63
00:03:41,000 --> 00:03:44,880
Speaker 1: process that lasted for a year, and in February two

64
00:03:44,920 --> 00:03:48,280
Speaker 1: thousand and twelve, this proposal was selected as the next

65
00:03:48,320 --> 00:03:52,520
Speaker 1: Man Challenge for IBM research and we started to walk

66
00:03:52,560 --> 00:03:56,440
Speaker 1: a few months later with a small team that gradually expanded,

67
00:03:57,480 --> 00:04:01,839
Speaker 1: and we walked on that intensively for I would say

68
00:04:01,880 --> 00:04:06,160
Speaker 1: six and a half yels dedicated solewly to dismission of

69
00:04:06,600 --> 00:04:09,880
Speaker 1: developing a machine that will be able to debate humans.

70
00:04:10,960 --> 00:04:15,840
Speaker 1: And eventually we demonstrated this system in a in a

71
00:04:15,880 --> 00:04:18,240
Speaker 1: full life debate. It was a little bit more than

72
00:04:18,279 --> 00:04:21,680
Speaker 1: a year ago, and it was a debate between this

73
00:04:21,760 --> 00:04:26,279
Speaker 1: system now being called the project debate and one of

74
00:04:26,400 --> 00:04:31,880
Speaker 1: the legendary debates in the history of university debate competitions,

75
00:04:32,000 --> 00:04:35,400
Speaker 1: and it still Harris Naam. It was in San Francisco,

76
00:04:35,680 --> 00:04:39,240
Speaker 1: and and it was a full life debate, surprisingly reminiscent

77
00:04:39,320 --> 00:04:42,640
Speaker 1: to division that we had back in the office in

78
00:04:42,640 --> 00:04:46,960
Speaker 1: Tel Aviv quite a few fields earlier in that single side.

79
00:04:47,360 --> 00:04:49,520
Speaker 1: So the topic of debate brings with it a few

80
00:04:49,520 --> 00:04:52,920
Speaker 1: different connotations, um, you know, and therefore the idea of

81
00:04:53,240 --> 00:04:56,000
Speaker 1: AI entering the frame might might be a bit confusing

82
00:04:56,080 --> 00:04:58,679
Speaker 1: for for some you know, we might imagine a computer

83
00:04:58,839 --> 00:05:01,840
Speaker 1: designed to defeat play or or perhaps a robot that

84
00:05:01,920 --> 00:05:06,279
Speaker 1: can shout louder and a televised US presidential debate to Daddy,

85
00:05:06,320 --> 00:05:09,360
Speaker 1: and can you walk us through what Project Debater is

86
00:05:09,600 --> 00:05:14,160
Speaker 1: and perhaps what it isn't. Yes, absolutely so. So first

87
00:05:14,160 --> 00:05:17,120
Speaker 1: of all, it is worth explaining what we mean, indeed

88
00:05:17,120 --> 00:05:21,800
Speaker 1: by a debate between an AI system like Project Debata

89
00:05:21,960 --> 00:05:26,960
Speaker 1: and a human opponent. So the debate starts with with

90
00:05:27,080 --> 00:05:30,640
Speaker 1: a motion in the debate jargon that defines what we're

91
00:05:30,680 --> 00:05:34,880
Speaker 1: going to debate. And in the event in San Francisco,

92
00:05:35,000 --> 00:05:38,080
Speaker 1: the topic was whether or not the government should subsidize

93
00:05:38,320 --> 00:05:42,159
Speaker 1: the schools. Uh. There are many considerations around how this

94
00:05:42,320 --> 00:05:44,839
Speaker 1: topic is being selected which we can skip, but the

95
00:05:44,920 --> 00:05:48,160
Speaker 1: only thing we should really emphasize is that this topic

96
00:05:49,080 --> 00:05:52,279
Speaker 1: is selected from a list of topics that were never

97
00:05:52,400 --> 00:05:57,039
Speaker 1: included in the training of the system, So the system

98
00:05:57,120 --> 00:06:00,160
Speaker 1: was never able to train on this particular topic. It

99
00:06:00,320 --> 00:06:03,880
Speaker 1: was trying to debate a new topic from from the

100
00:06:04,240 --> 00:06:08,000
Speaker 1: perspective of the machine. And then we are on the

101
00:06:08,000 --> 00:06:10,680
Speaker 1: side of the governments of Project Debta is supporting the

102
00:06:10,760 --> 00:06:14,200
Speaker 1: motion and how the issues on the opposition, and we

103
00:06:14,320 --> 00:06:18,039
Speaker 1: have a full minutes opening speeches for each side and

104
00:06:18,160 --> 00:06:23,480
Speaker 1: full minutely bottom speeches and two minutes closing statements. So

105
00:06:23,520 --> 00:06:26,400
Speaker 1: all you know, we are talking about a little more

106
00:06:26,440 --> 00:06:29,760
Speaker 1: than twenty to twenty five minutes of a discussion that

107
00:06:29,880 --> 00:06:34,800
Speaker 1: we hope we will be a meaningful discussion between Project

108
00:06:34,839 --> 00:06:38,000
Speaker 1: Debata and and and a human plish in these particularly

109
00:06:38,760 --> 00:06:42,240
Speaker 1: so to clarify for people who might not be familiar

110
00:06:42,279 --> 00:06:45,920
Speaker 1: with competitive debating. So competitive debating does not involve what

111
00:06:46,080 --> 00:06:48,640
Speaker 1: people might be more familiar with, which is like passionately

112
00:06:48,760 --> 00:06:52,279
Speaker 1: arguing your actual point of view. It involves having a

113
00:06:52,360 --> 00:06:56,080
Speaker 1: position selected for you that you then must get up

114
00:06:56,120 --> 00:06:59,440
Speaker 1: and defend in front of the judges. Correct, yes, this

115
00:06:59,520 --> 00:07:02,320
Speaker 1: is called act and and this is indeed important to

116
00:07:02,360 --> 00:07:06,719
Speaker 1: emphasize because you do not know in advance what is

117
00:07:06,720 --> 00:07:10,440
Speaker 1: going to be your side. And and even if you

118
00:07:10,480 --> 00:07:12,080
Speaker 1: know in advance that you are going to be on

119
00:07:12,120 --> 00:07:15,440
Speaker 1: the side of the government, we should bear in mind

120
00:07:15,440 --> 00:07:19,360
Speaker 1: the motion could have been phrased we should not subsidize previously,

121
00:07:20,280 --> 00:07:23,640
Speaker 1: and then you should actually contest that. So you do

122
00:07:23,720 --> 00:07:25,800
Speaker 1: not know in advance what is going to be your

123
00:07:25,880 --> 00:07:29,080
Speaker 1: stance to the topic. This is true for Project Debata

124
00:07:29,160 --> 00:07:32,960
Speaker 1: and also for the for the human opponent, and you

125
00:07:33,080 --> 00:07:36,240
Speaker 1: have only ten to fifteen minutes to PerPell. You don't

126
00:07:36,240 --> 00:07:38,680
Speaker 1: know the topic in advance. This is again true for

127
00:07:39,160 --> 00:07:44,240
Speaker 1: project debata and for the human opponent, and uh, your

128
00:07:44,320 --> 00:07:47,280
Speaker 1: goal is really to to persuade the audience. And this

129
00:07:47,440 --> 00:07:50,960
Speaker 1: actually touches on an interesting question of how do you

130
00:07:51,200 --> 00:07:56,240
Speaker 1: do you measure who won the debate? Because in chess

131
00:07:56,320 --> 00:07:59,040
Speaker 1: and in other games this is very clear and and

132
00:07:59,240 --> 00:08:03,280
Speaker 1: really part of the problem with with with debate in

133
00:08:03,320 --> 00:08:07,840
Speaker 1: general and with developing artificial intelligence that is capable of

134
00:08:07,880 --> 00:08:11,640
Speaker 1: debating in particular now is that it is very hard

135
00:08:12,040 --> 00:08:15,160
Speaker 1: to to be fine who actually won the debate. Yeah,

136
00:08:15,200 --> 00:08:17,680
Speaker 1: I know. There are a couple of different metrics. So

137
00:08:17,760 --> 00:08:20,200
Speaker 1: of course one would just be like, what is the

138
00:08:20,360 --> 00:08:23,320
Speaker 1: percentage of the audience that is convinced to either side?

139
00:08:23,320 --> 00:08:25,560
Speaker 1: But that can be problematic because people come in with

140
00:08:25,600 --> 00:08:28,800
Speaker 1: their own opinions already formed on an issue. So one

141
00:08:29,680 --> 00:08:33,480
Speaker 1: metric I've seen is how much the percentages change. They

142
00:08:33,559 --> 00:08:37,920
Speaker 1: ask people before and afterward what their positions are, and

143
00:08:37,960 --> 00:08:41,680
Speaker 1: then after word they say, okay, which side has one

144
00:08:41,840 --> 00:08:45,199
Speaker 1: over more people? Whatever the starting percentages were is, And

145
00:08:45,440 --> 00:08:48,600
Speaker 1: I assume you all had a metric like that precisely so,

146
00:08:48,600 --> 00:08:51,800
Speaker 1: so this is exactly the point, because if you simply

147
00:08:51,840 --> 00:08:54,520
Speaker 1: ask people who is more convinced, you need somehow to

148
00:08:54,559 --> 00:08:58,600
Speaker 1: take into account the opinions to begin with, and and

149
00:08:58,640 --> 00:09:01,800
Speaker 1: the it is done exactly as as you described it.

150
00:09:01,880 --> 00:09:06,559
Speaker 1: And all this event was in collaboration with with Intelligence

151
00:09:06,720 --> 00:09:10,000
Speaker 1: as well, which is really I think the leading platform

152
00:09:10,160 --> 00:09:14,640
Speaker 1: in the US for organizing such a high profile competitive debate.

153
00:09:15,480 --> 00:09:19,719
Speaker 1: It was hosted, the moderator was the moderator Form Intelligence

154
00:09:19,800 --> 00:09:23,480
Speaker 1: as well, John Dunvan, and and the voting was done

155
00:09:23,520 --> 00:09:27,840
Speaker 1: exactly as you described and as being done with the

156
00:09:27,880 --> 00:09:30,760
Speaker 1: show of Intelligence Square. That is, the audience is voting

157
00:09:31,240 --> 00:09:35,400
Speaker 1: before the debate starts, and they vote again after the

158
00:09:35,440 --> 00:09:38,680
Speaker 1: debate ends, and you win if you were able to

159
00:09:38,720 --> 00:09:41,640
Speaker 1: move more people to to your side. Now I think

160
00:09:41,640 --> 00:09:43,720
Speaker 1: a lot of people might be wondering, how on earth

161
00:09:43,720 --> 00:09:47,640
Speaker 1: would you even begin to organize a persuasive argument from

162
00:09:47,679 --> 00:09:49,680
Speaker 1: an AI point of view? Could you walk us through

163
00:09:49,720 --> 00:09:53,960
Speaker 1: the technical specifics of how Project Debater would put together

164
00:09:54,040 --> 00:09:57,280
Speaker 1: an argument. Yes, so we were asking ourselves the same

165
00:09:57,400 --> 00:10:03,240
Speaker 1: question actually when when we started this project. And I

166
00:10:03,280 --> 00:10:07,480
Speaker 1: think this is part of the of the nature of

167
00:10:07,640 --> 00:10:10,920
Speaker 1: such a grand challenge that you do not really know

168
00:10:11,120 --> 00:10:15,800
Speaker 1: how exactly you are going to to approach the problem.

169
00:10:15,840 --> 00:10:21,000
Speaker 1: But we did what computer scientists often do, and this

170
00:10:21,080 --> 00:10:25,480
Speaker 1: is to take this big and somewhat amorphic problem and

171
00:10:25,600 --> 00:10:31,320
Speaker 1: break it into more modular and hopefully more tangible tasks.

172
00:10:31,480 --> 00:10:37,640
Speaker 1: And so in general, the debated system had uh two

173
00:10:37,760 --> 00:10:41,520
Speaker 1: major sources of information. One of them is the massive

174
00:10:41,600 --> 00:10:48,480
Speaker 1: collection of around four hundred million newspaper articles, and when

175
00:10:48,480 --> 00:10:54,679
Speaker 1: the debate starts, the system was using various AI artificial

176
00:10:54,720 --> 00:11:00,840
Speaker 1: intelligence engines in order to try and pinpoint short pieces

177
00:11:00,880 --> 00:11:04,480
Speaker 1: of text within this massive collection. We're talking about ten

178
00:11:04,600 --> 00:11:09,120
Speaker 1: billion sentences, so we were trying to automatically pinpoint these

179
00:11:09,200 --> 00:11:14,880
Speaker 1: short pieces of text that should satisfy three criteria. They

180
00:11:14,880 --> 00:11:19,200
Speaker 1: should be relevant to the topic, they should be argumentative

181
00:11:19,240 --> 00:11:22,880
Speaker 1: in nature, they should argue something about the topic, and

182
00:11:22,920 --> 00:11:26,720
Speaker 1: they should support our side of the debate. And this

183
00:11:26,800 --> 00:11:30,280
Speaker 1: is quite a formidable challenge. But assuming that you are

184
00:11:30,320 --> 00:11:33,360
Speaker 1: capable of finding these short pieces of tax, the system

185
00:11:33,440 --> 00:11:38,040
Speaker 1: is then using other AI capabilities in order to try

186
00:11:38,160 --> 00:11:42,559
Speaker 1: and glue them together into a meaningful narrative. So this

187
00:11:42,679 --> 00:11:47,080
Speaker 1: is one major source of information for the system. The

188
00:11:47,160 --> 00:11:50,600
Speaker 1: second important source of information for the system was a

189
00:11:50,800 --> 00:11:58,480
Speaker 1: unique collection of more principled arguments that were actually written

190
00:11:58,800 --> 00:12:03,320
Speaker 1: by by humans, and we are talking about thousands of

191
00:12:03,480 --> 00:12:07,400
Speaker 1: more principled arguments. And the role of the system was

192
00:12:07,440 --> 00:12:11,080
Speaker 1: when the debate starts, was really to navigate within this

193
00:12:11,200 --> 00:12:14,880
Speaker 1: collection and find the most relevant principled arguments and use

194
00:12:14,960 --> 00:12:17,280
Speaker 1: them in the right timing. So so, to make this

195
00:12:17,360 --> 00:12:21,839
Speaker 1: more concrete what we mean by a principal argument, imagine

196
00:12:21,880 --> 00:12:25,200
Speaker 1: that we are debating whether or not to ban organ

197
00:12:25,280 --> 00:12:28,400
Speaker 1: trade or whether or not to ban the sale of alcohol.

198
00:12:28,880 --> 00:12:32,040
Speaker 1: In both cases, the opposition may argue that if you

199
00:12:32,120 --> 00:12:35,520
Speaker 1: ban something, you are at the risk of the emergence

200
00:12:35,559 --> 00:12:38,120
Speaker 1: of a black market. So a black market is a

201
00:12:38,160 --> 00:12:41,280
Speaker 1: principled argument that can be used almost in the same

202
00:12:41,360 --> 00:12:46,400
Speaker 1: way in many different contexts. So one may naively assume

203
00:12:47,360 --> 00:12:51,760
Speaker 1: that this is kind of a simple keyword matching thing.

204
00:12:52,040 --> 00:12:55,560
Speaker 1: If we ban something, then the opposition is going to

205
00:12:55,679 --> 00:12:58,240
Speaker 1: use the black market argument, and we should be prepared

206
00:12:58,320 --> 00:13:02,000
Speaker 1: for that. But obviously this is far from true. So

207
00:13:02,120 --> 00:13:07,679
Speaker 1: imagine a debate about banning breastfeeding in public. Obviously there

208
00:13:07,800 --> 00:13:11,400
Speaker 1: is little risk for a black market in this contract.

209
00:13:11,520 --> 00:13:15,000
Speaker 1: Or imagine a debate about banning internet cookies. We're not

210
00:13:15,120 --> 00:13:18,320
Speaker 1: going to tee a black market of internet cookies if

211
00:13:18,360 --> 00:13:22,560
Speaker 1: we band these. So the system really needs to develop

212
00:13:22,640 --> 00:13:27,960
Speaker 1: a more subtle understanding after human language in order to

213
00:13:28,080 --> 00:13:31,920
Speaker 1: be able to identify the most relevant principle argument and

214
00:13:31,960 --> 00:13:35,400
Speaker 1: need use them doing a debate. And and this is,

215
00:13:35,440 --> 00:13:40,199
Speaker 1: by the way, just what all this description is before

216
00:13:40,320 --> 00:13:43,800
Speaker 1: listening to the opponent. This is just what we're going

217
00:13:43,880 --> 00:13:47,840
Speaker 1: to say on our side. And and the most the

218
00:13:47,920 --> 00:13:51,960
Speaker 1: most challenging part is really too uh to listen to

219
00:13:52,000 --> 00:13:55,199
Speaker 1: the opponent. And it's some kind of a battle to

220
00:13:55,320 --> 00:13:59,040
Speaker 1: the arguments generated by the opponment raised by the And

221
00:13:59,200 --> 00:14:03,920
Speaker 1: we do that you using uh an arsenal of technique

222
00:14:04,320 --> 00:14:07,240
Speaker 1: that most of them rely on the same principle. We

223
00:14:07,360 --> 00:14:11,800
Speaker 1: start by listening to the world articulated by the opponment,

224
00:14:11,920 --> 00:14:14,760
Speaker 1: and for that we simply use what's on speech recognition

225
00:14:15,200 --> 00:14:17,839
Speaker 1: capabilities out of the box. But of course we need

226
00:14:17,880 --> 00:14:20,360
Speaker 1: to go to beyond the world, and we need to

227
00:14:20,440 --> 00:14:23,520
Speaker 1: understand the gist of the arguments of the opponent. And

228
00:14:23,560 --> 00:14:27,640
Speaker 1: in order to do that we try using various smackloads

229
00:14:27,680 --> 00:14:33,200
Speaker 1: to anticipate in advance what kind of arguments the opposition

230
00:14:33,800 --> 00:14:38,520
Speaker 1: mind you and then listen to determine whether he did

231
00:14:38,560 --> 00:14:43,720
Speaker 1: the opposition was making these arguments and then responded cold yeah.

232
00:14:43,800 --> 00:14:47,920
Speaker 1: That calls to mind the question of the difference between, say,

233
00:14:47,960 --> 00:14:51,720
Speaker 1: what's a sound argument versus what's a persuasive argument? I mean,

234
00:14:51,960 --> 00:14:55,840
Speaker 1: we know from reality that often the most persuasive appeals

235
00:14:55,840 --> 00:15:00,680
Speaker 1: and debates rely on just straightforwardly false claims and logical fallacies,

236
00:15:00,840 --> 00:15:03,960
Speaker 1: or even on little emotional cues that have little to

237
00:15:04,000 --> 00:15:06,680
Speaker 1: do with the matter at hand. I was thinking about

238
00:15:06,680 --> 00:15:09,240
Speaker 1: how in live debates, if you can get a laugh

239
00:15:09,400 --> 00:15:12,560
Speaker 1: at your opponent's expense, that's worth you know, a dozen

240
00:15:13,640 --> 00:15:18,200
Speaker 1: sound arguments or claims. So to what degree can AI

241
00:15:18,360 --> 00:15:21,840
Speaker 1: understand these sorts of persuasive appeals that that go beyond

242
00:15:22,000 --> 00:15:24,560
Speaker 1: just like what kind of evidence you can bring and

243
00:15:24,640 --> 00:15:29,880
Speaker 1: the appeals based on style you're right in in in

244
00:15:29,880 --> 00:15:33,040
Speaker 1: in debate and in the methods. We know already from

245
00:15:33,080 --> 00:15:38,320
Speaker 1: the ancient weeks that that we have free elaps, we

246
00:15:38,440 --> 00:15:43,240
Speaker 1: have logos, and we have ethos, and we have afforts,

247
00:15:43,280 --> 00:15:47,120
Speaker 1: and humans are using a mixture of these pilas when

248
00:15:47,440 --> 00:15:51,600
Speaker 1: they are debating one another. And just as a quick clarification, logos,

249
00:15:51,640 --> 00:15:55,080
Speaker 1: pathos and ethos are the types of appeals that were

250
00:15:55,120 --> 00:15:58,840
Speaker 1: identified in the study of classical rhetoric. Where logos is

251
00:15:58,920 --> 00:16:03,160
Speaker 1: appeals based on our logical arguments and evidence, Pathos is

252
00:16:03,160 --> 00:16:06,200
Speaker 1: the appeal to the emotions or the passions, and ethos

253
00:16:06,320 --> 00:16:09,440
Speaker 1: is an appeal based on the credibility or authority of

254
00:16:09,480 --> 00:16:14,360
Speaker 1: the speaker. I mean, as you know broadly understood and

255
00:16:14,360 --> 00:16:18,800
Speaker 1: and the technology that we developed, and and by the way,

256
00:16:18,800 --> 00:16:23,600
Speaker 1: it should be stated that there is a rapidly emerging

257
00:16:23,680 --> 00:16:28,880
Speaker 1: community of scientists across the globe that are investigating this

258
00:16:29,080 --> 00:16:32,120
Speaker 1: kind of topic. It is all under the umbrella of

259
00:16:32,240 --> 00:16:38,240
Speaker 1: this emerging field, yeah, referred to as a computational argumentation.

260
00:16:38,960 --> 00:16:41,760
Speaker 1: And when we started in two thousand and twelve, there

261
00:16:41,920 --> 00:16:46,720
Speaker 1: was a handful of teams pursuing that, and we see

262
00:16:46,720 --> 00:16:50,160
Speaker 1: a very dramatic increase in the result in these areas

263
00:16:50,200 --> 00:16:54,160
Speaker 1: of the last few years is very I think from

264
00:16:55,000 --> 00:17:01,880
Speaker 1: exacting and as I mentioned, the technology that we developed

265
00:17:01,880 --> 00:17:06,159
Speaker 1: a most focused on logos, and you can see in

266
00:17:06,200 --> 00:17:09,640
Speaker 1: the debate between proper Debate and Hali. By the way,

267
00:17:09,680 --> 00:17:13,879
Speaker 1: this this debate is is fully available on YouTube, and

268
00:17:14,080 --> 00:17:19,080
Speaker 1: you can see that indeed a woman is better in

269
00:17:19,240 --> 00:17:23,560
Speaker 1: making in using path as and perhaps in using ethos

270
00:17:23,600 --> 00:17:27,040
Speaker 1: and it is harder for the machine. And indeed most

271
00:17:27,040 --> 00:17:30,800
Speaker 1: of the research being done by by the by the

272
00:17:30,880 --> 00:17:35,560
Speaker 1: relevant research communities around logos, but there are already attempt

273
00:17:36,040 --> 00:17:40,320
Speaker 1: trying to model and to capture additional aspect of path

274
00:17:40,359 --> 00:17:44,400
Speaker 1: of and ethos in all the further enhanced this kind

275
00:17:44,400 --> 00:17:48,240
Speaker 1: of technology. So another question I have is debater has

276
00:17:48,280 --> 00:17:52,879
Speaker 1: to source claims and facts and arguments from existing written

277
00:17:52,880 --> 00:17:55,400
Speaker 1: work produced by humans, which of course we know can

278
00:17:55,440 --> 00:17:58,280
Speaker 1: be full of all sorts of flaws. Is there any

279
00:17:58,280 --> 00:18:01,600
Speaker 1: way at this point for it to to have an

280
00:18:01,600 --> 00:18:05,960
Speaker 1: analytical function to tell a say, factually true claim or

281
00:18:06,000 --> 00:18:09,960
Speaker 1: a logically valid argument from just something that is wrong

282
00:18:10,080 --> 00:18:12,920
Speaker 1: or dubious but repeated a lot in writing, or are

283
00:18:13,000 --> 00:18:19,200
Speaker 1: we not there yet? This is a very kindly important

284
00:18:19,240 --> 00:18:24,000
Speaker 1: and difficult problem, and that is receiving going attempting over

285
00:18:24,280 --> 00:18:30,639
Speaker 1: over the previous teams and go to tackle that. This

286
00:18:30,840 --> 00:18:35,080
Speaker 1: is certainly not bullet bof and and the problem is

287
00:18:35,080 --> 00:18:39,520
Speaker 1: is quite complex because one may say, you know, okay, fine,

288
00:18:39,600 --> 00:18:43,600
Speaker 1: maybe I should only take my argument from highly credibally

289
00:18:43,640 --> 00:18:49,760
Speaker 1: so and by boxy I can assume that that these

290
00:18:49,880 --> 00:18:54,720
Speaker 1: arguments are our valid. But this is not necessarily the case. Right.

291
00:18:54,800 --> 00:18:58,240
Speaker 1: You can see you can lead an opinion article in

292
00:18:58,359 --> 00:19:05,240
Speaker 1: a highly respectable newspaper which is actually quoting a false

293
00:19:05,359 --> 00:19:08,879
Speaker 1: argument that was made as well, and if you're not

294
00:19:08,920 --> 00:19:13,440
Speaker 1: careful enough, you you might be your system is going

295
00:19:13,480 --> 00:19:17,440
Speaker 1: to pull this argument without understanding that something is happening.

296
00:19:18,119 --> 00:19:21,199
Speaker 1: So we try to develop and we actually part of

297
00:19:21,240 --> 00:19:26,800
Speaker 1: Project Debate included some kind of filtering mechanism in order

298
00:19:26,920 --> 00:19:30,119
Speaker 1: to to filter out these kind of cases. And the

299
00:19:30,200 --> 00:19:34,359
Speaker 1: way we did that was really once a specific claim

300
00:19:34,920 --> 00:19:37,800
Speaker 1: was affected and by the way to being ordered, the

301
00:19:37,960 --> 00:19:41,200
Speaker 1: claim is not a full sentence. A claim is often

302
00:19:41,600 --> 00:19:44,600
Speaker 1: only a part of a tentence. Even if you were

303
00:19:44,680 --> 00:19:48,720
Speaker 1: able to detect sentence that contains a claim relevant one

304
00:19:48,800 --> 00:19:51,800
Speaker 1: that supportal side out of the billions of sentences in

305
00:19:51,840 --> 00:19:55,160
Speaker 1: the popos, you still need to find the coret boundaries

306
00:19:55,560 --> 00:19:58,520
Speaker 1: after claim within the sentence, and you have hundreds of

307
00:19:58,600 --> 00:20:02,440
Speaker 1: options and only all of them is correct. So this

308
00:20:02,520 --> 00:20:05,320
Speaker 1: is just going back why this this problem is it

309
00:20:05,440 --> 00:20:08,760
Speaker 1: so talenting? But until you do that and found this

310
00:20:08,960 --> 00:20:12,119
Speaker 1: claim and asked what is the stance of this claim,

311
00:20:12,440 --> 00:20:15,560
Speaker 1: and if the stance is supporting your side, you can

312
00:20:15,600 --> 00:20:18,920
Speaker 1: still ask what is the stance of the full sentence?

313
00:20:20,200 --> 00:20:22,439
Speaker 1: And if the stance of the full sentences in the

314
00:20:22,480 --> 00:20:26,520
Speaker 1: opposite direction, you may suspect that something is going on.

315
00:20:27,359 --> 00:20:31,280
Speaker 1: And perhaps this this claim is quoted in order to

316
00:20:31,920 --> 00:20:35,680
Speaker 1: contradict and not because it is true. And then perhaps

317
00:20:35,680 --> 00:20:39,879
Speaker 1: it is there it is safer to avoid using it.

318
00:20:39,920 --> 00:20:44,120
Speaker 1: But but this is just one safety mechanism, and and

319
00:20:44,200 --> 00:20:46,880
Speaker 1: the problem that you raise is actually a much more

320
00:20:46,960 --> 00:20:52,080
Speaker 1: beneval one, and and I think many teams are working

321
00:20:52,119 --> 00:20:55,359
Speaker 1: on that, and we try to address that as well.

322
00:20:55,600 --> 00:21:00,280
Speaker 1: And I think it has many interesting dimensions because it

323
00:21:00,400 --> 00:21:04,600
Speaker 1: is not even just about the validity of the argument. Often,

324
00:21:04,680 --> 00:21:08,600
Speaker 1: when when you show people to arguments, they will agree

325
00:21:08,640 --> 00:21:11,720
Speaker 1: that one of them is better than the other. But

326
00:21:11,920 --> 00:21:15,920
Speaker 1: what are the underlying mechanisms that I'd ask to the

327
00:21:16,160 --> 00:21:19,439
Speaker 1: one argument over the other, And how do you train

328
00:21:19,800 --> 00:21:23,639
Speaker 1: an artificial as in system to make the distinction. This

329
00:21:23,840 --> 00:21:27,240
Speaker 1: is kind of another example of the problems that welcome

330
00:21:27,280 --> 00:21:30,440
Speaker 1: to them. I have a question about what could come

331
00:21:30,520 --> 00:21:33,960
Speaker 1: out of AI research like this, because I would say,

332
00:21:33,960 --> 00:21:36,879
Speaker 1: from my personal perspective, I think studying rhetoric and debate

333
00:21:37,040 --> 00:21:43,040
Speaker 1: is extremely important, but not necessarily because getting into debates

334
00:21:43,160 --> 00:21:45,760
Speaker 1: is a good way to figure out what's true and

335
00:21:45,920 --> 00:21:48,040
Speaker 1: establish you know, the right thing to do. I think

336
00:21:48,119 --> 00:21:50,720
Speaker 1: one of the most important reasons to study rhetoric and

337
00:21:50,760 --> 00:21:54,800
Speaker 1: debate is so that you can understand how other people's

338
00:21:54,960 --> 00:21:58,760
Speaker 1: arguments and persuasive appeals are operating on you, or are

339
00:21:58,880 --> 00:22:02,280
Speaker 1: designed to operate you. A clear understanding of rhetoric can

340
00:22:02,280 --> 00:22:04,959
Speaker 1: be a kind of suit of armor for going into

341
00:22:05,160 --> 00:22:08,879
Speaker 1: you know, the world and seeing how political actors and

342
00:22:08,960 --> 00:22:11,960
Speaker 1: business actors and advertising and all that is trying to

343
00:22:12,040 --> 00:22:15,720
Speaker 1: affect you. Do you see project debate or serving any

344
00:22:15,800 --> 00:22:19,080
Speaker 1: kind of educational purpose like this in the world today.

345
00:22:19,560 --> 00:22:25,080
Speaker 1: So there are several levels by which I can I

346
00:22:25,119 --> 00:22:30,680
Speaker 1: can answer that. The first one is that this kind

347
00:22:30,680 --> 00:22:36,320
Speaker 1: of technology is is definitely relevant and we believe highly

348
00:22:36,520 --> 00:22:42,200
Speaker 1: valuable in the context of education. You can imagine using

349
00:22:42,240 --> 00:22:47,080
Speaker 1: the technology in order to build better arguments and more

350
00:22:47,119 --> 00:22:53,640
Speaker 1: of all, to perform a more analytical and perhaps more

351
00:22:53,680 --> 00:23:01,640
Speaker 1: objective analysis off complex and controversial topics. This is one aspect.

352
00:23:02,560 --> 00:23:06,280
Speaker 1: There is another aspect, but often when we debate is

353
00:23:06,920 --> 00:23:13,639
Speaker 1: other humans. There are many layouts that that are involved

354
00:23:13,840 --> 00:23:16,560
Speaker 1: in this discussion. In this debate. What all of them

355
00:23:16,560 --> 00:23:20,119
Speaker 1: are related? To the facts and to the arguments that

356
00:23:20,200 --> 00:23:23,400
Speaker 1: we are raising. Perhaps we have history with that Belton,

357
00:23:23,880 --> 00:23:28,040
Speaker 1: Perhaps we have history with ourselves that actually impact our

358
00:23:28,119 --> 00:23:32,600
Speaker 1: on part and decisions. Perhaps other people are listening and

359
00:23:32,680 --> 00:23:38,480
Speaker 1: this actually improvides contact, uh that impact what is happening.

360
00:23:38,960 --> 00:23:42,720
Speaker 1: And we are curious about this option of the dating

361
00:23:42,800 --> 00:23:47,680
Speaker 1: with the machine in the privacy of your office. Maybe

362
00:23:47,720 --> 00:23:51,840
Speaker 1: this is a different form of a discussion that to

363
00:23:52,000 --> 00:23:58,159
Speaker 1: some extent is perhaps all free of of external biases

364
00:23:58,240 --> 00:24:04,280
Speaker 1: and maybe will enable treat some people to identify situations

365
00:24:04,280 --> 00:24:08,000
Speaker 1: where they have a blind book and to better listen

366
00:24:08,320 --> 00:24:12,159
Speaker 1: to the other side. So I think in this case

367
00:24:12,280 --> 00:24:17,400
Speaker 1: the whole of the technology could be quite instrumental and positive.

368
00:24:17,840 --> 00:24:21,880
Speaker 1: The false business applications that are also very interesting from

369
00:24:21,920 --> 00:24:28,240
Speaker 1: the IBM perspective and uh, and this is another another dimension,

370
00:24:28,520 --> 00:24:37,280
Speaker 1: another level by which we can consider the technology as exacuable. Again,

371
00:24:37,280 --> 00:24:39,720
Speaker 1: big thanks to No One slow name for taking time

372
00:24:39,760 --> 00:24:41,600
Speaker 1: to chat with us. And now we're going to go

373
00:24:41,640 --> 00:24:49,240
Speaker 1: straight into our second talk on the subject with Madu Matt.

374
00:24:49,520 --> 00:24:51,960
Speaker 1: Thanks so much for joining us today. Could you start

375
00:24:52,000 --> 00:24:56,320
Speaker 1: off by introducing yourself and talking about your role at IBM. Yeah, absolutely,

376
00:24:56,440 --> 00:24:59,920
Speaker 1: and really nice to meet you. Robert and Joe uh

377
00:25:00,040 --> 00:25:06,040
Speaker 1: maduco Chi, vice President Offering Management in Data and AI IBM,

378
00:25:06,080 --> 00:25:09,159
Speaker 1: And the role of offering management is really all about

379
00:25:09,680 --> 00:25:14,040
Speaker 1: laying down the strategy and then delivering and executing towards

380
00:25:14,119 --> 00:25:18,520
Speaker 1: such strategy. And I'm based out of San Jose, Sunny, California, excellent.

381
00:25:19,359 --> 00:25:21,679
Speaker 1: So just to kick things off here, um, you know

382
00:25:21,680 --> 00:25:24,600
Speaker 1: we're gonna be talking a lot about AI here, and

383
00:25:25,440 --> 00:25:28,280
Speaker 1: it makes sense to to to really get into what

384
00:25:28,359 --> 00:25:31,520
Speaker 1: we mean when we're talking about AI for business. How

385
00:25:31,560 --> 00:25:35,560
Speaker 1: does AI serve business compared to the way it serves consumers.

386
00:25:36,359 --> 00:25:39,520
Speaker 1: That's a great question to get started on. UM so

387
00:25:40,160 --> 00:25:44,679
Speaker 1: redeveloped a thesis a couple of years ago about really

388
00:25:44,720 --> 00:25:50,000
Speaker 1: how AI for business would be different from consumer AI.

389
00:25:50,280 --> 00:25:53,159
Speaker 1: Think of consumer AI, which we all know work with

390
00:25:53,200 --> 00:25:58,200
Speaker 1: our smartphones, smart speakers, social media, photos, everything what it comes.

391
00:25:58,280 --> 00:26:01,440
Speaker 1: But when it comes for AI for business, it's really

392
00:26:01,560 --> 00:26:06,880
Speaker 1: very very different. AI for business is all about automation,

393
00:26:07,280 --> 00:26:12,560
Speaker 1: optimization and making better predictions, and it requires really a

394
00:26:12,680 --> 00:26:16,000
Speaker 1: very different set of technical capabilities, like you would have

395
00:26:16,040 --> 00:26:19,080
Speaker 1: to understand how to deal with language, have to deal

396
00:26:19,119 --> 00:26:23,240
Speaker 1: with what does automation means, and then be able to

397
00:26:23,680 --> 00:26:28,120
Speaker 1: have the explainability and trust up AI. UM. So that's

398
00:26:28,119 --> 00:26:31,200
Speaker 1: sort of the big difference between commercial AI and AI

399
00:26:31,280 --> 00:26:33,560
Speaker 1: for business. So we know that one of the big

400
00:26:33,600 --> 00:26:36,280
Speaker 1: AI projects at IBM is Watson. Could you tell us

401
00:26:36,320 --> 00:26:39,880
Speaker 1: about Watson and explain how Watson fits into the broader

402
00:26:39,920 --> 00:26:44,960
Speaker 1: picture of recent advancements in AI. Sure you you might

403
00:26:45,000 --> 00:26:47,760
Speaker 1: have heard of Watson, and our audience might have heard

404
00:26:47,800 --> 00:26:50,800
Speaker 1: of Watson, which came out when we first did our

405
00:26:52,000 --> 00:26:56,239
Speaker 1: UH in Jeopardy and people remember Watson from there. But

406
00:26:56,680 --> 00:26:59,840
Speaker 1: fast forward, a lot of work done around Watson. Think

407
00:26:59,840 --> 00:27:05,160
Speaker 1: of Watson as our definition of IBM AI. We evolved

408
00:27:05,200 --> 00:27:10,120
Speaker 1: a lot um since then, and our strategic intent always

409
00:27:10,160 --> 00:27:15,240
Speaker 1: has been to have what's an available anywhere meaning available

410
00:27:15,359 --> 00:27:20,600
Speaker 1: on any cloud. UH. We have focused on Watson. UH

411
00:27:21,080 --> 00:27:24,400
Speaker 1: we call with Watson meaning it's embedded in almost all

412
00:27:24,440 --> 00:27:29,080
Speaker 1: your applications. So for example, UM, I use the world

413
00:27:29,119 --> 00:27:32,159
Speaker 1: a lot for AI for AI. What does that mean? Like,

414
00:27:32,280 --> 00:27:35,879
Speaker 1: how do we embed AI in our data sciences and

415
00:27:35,960 --> 00:27:40,840
Speaker 1: in our data data platforms and such. The other parts

416
00:27:40,880 --> 00:27:44,480
Speaker 1: of evolution has been you know, as I said earlier,

417
00:27:44,720 --> 00:27:48,320
Speaker 1: from our AI for business is all about automation. How

418
00:27:48,359 --> 00:27:52,960
Speaker 1: do we UH evolve into the workflow AI that matters

419
00:27:53,359 --> 00:27:57,960
Speaker 1: for our clients and our our society. So the workflows

420
00:27:58,280 --> 00:28:02,439
Speaker 1: could definition could be you know, customer care, uh in

421
00:28:02,680 --> 00:28:07,600
Speaker 1: I t asset management, in your regulatory or compliance, in

422
00:28:07,760 --> 00:28:11,960
Speaker 1: supply chain or in your planning and budgeting. Right, these

423
00:28:12,000 --> 00:28:15,679
Speaker 1: are how you can really embed AI and that is

424
00:28:15,720 --> 00:28:20,720
Speaker 1: where Watson has really evolved into. And we have also

425
00:28:20,960 --> 00:28:25,840
Speaker 1: been delivering now Watson an AI capability in a in

426
00:28:25,880 --> 00:28:30,520
Speaker 1: our integrated single platform we call cloud Pacer data. So

427
00:28:30,880 --> 00:28:33,480
Speaker 1: a long way. We came from Jeopardy Days and then

428
00:28:33,560 --> 00:28:36,359
Speaker 1: you just heard from nome where we landed with Debater.

429
00:28:36,920 --> 00:28:42,000
Speaker 1: So speaking of Debater, what capabilities has IBM commercialized from

430
00:28:42,080 --> 00:28:46,840
Speaker 1: Project Debater into Watson? So that's a great question. Um,

431
00:28:47,000 --> 00:28:52,200
Speaker 1: A lot of commercialization has happened. We have uh pretty

432
00:28:52,240 --> 00:28:56,080
Speaker 1: good rich set of products like what's an assistant, what's

433
00:28:56,120 --> 00:28:59,800
Speaker 1: on discovery, what's on knowledge, language understanding? And I know

434
00:28:59,840 --> 00:29:02,120
Speaker 1: the are just works, but let me just give a

435
00:29:03,040 --> 00:29:05,960
Speaker 1: bit of a background on what what's an assistant is?

436
00:29:06,040 --> 00:29:11,560
Speaker 1: What's an assistant is? Our conversational AI platform really helps

437
00:29:11,600 --> 00:29:17,800
Speaker 1: provide customer fast, straightforward answer, accurate answers UM across any application,

438
00:29:17,960 --> 00:29:21,720
Speaker 1: device or cloud right, UM and our discovery is all

439
00:29:21,760 --> 00:29:27,480
Speaker 1: about enterprise search and AI search technology that truly retrieves

440
00:29:27,720 --> 00:29:32,400
Speaker 1: specific answers to your questions while you're analyzing trends and

441
00:29:32,480 --> 00:29:36,760
Speaker 1: relationships in the enterprise data. So we've been looking at

442
00:29:36,840 --> 00:29:39,320
Speaker 1: debater and some of the key technologies. Let me give

443
00:29:39,320 --> 00:29:44,840
Speaker 1: you an example of few UM like sentiment analysis. Uh,

444
00:29:45,120 --> 00:29:48,200
Speaker 1: let me pose a problem statement, what does that really mean? So,

445
00:29:48,320 --> 00:29:54,600
Speaker 1: for example, today Watson does not understand idioms or sentiment shifters,

446
00:29:54,600 --> 00:29:58,040
Speaker 1: and neither does any other competitor operates out there also,

447
00:29:58,520 --> 00:30:04,040
Speaker 1: So think of elements which include hardly helpful, over the moon,

448
00:30:04,640 --> 00:30:08,000
Speaker 1: cold feet, UM all years. You know, how do you

449
00:30:08,080 --> 00:30:12,040
Speaker 1: make that analysis and figure figure this out? What is

450
00:30:12,080 --> 00:30:15,520
Speaker 1: the real context behind this? So what we have done

451
00:30:15,720 --> 00:30:19,000
Speaker 1: with that is that now what's on leverages this debat

452
00:30:19,160 --> 00:30:24,760
Speaker 1: technology and looks at these idioms and sentiment shifters and

453
00:30:24,840 --> 00:30:28,840
Speaker 1: does the analysis starting with better understanding of this sentiment

454
00:30:28,880 --> 00:30:31,840
Speaker 1: and analysis is one of the most widely used API

455
00:30:31,920 --> 00:30:35,400
Speaker 1: s for us UM. This already exists today in our

456
00:30:35,480 --> 00:30:40,520
Speaker 1: product portfolio. What's coming into the future is UM. It's

457
00:30:40,600 --> 00:30:44,320
Speaker 1: around all around documents. So let me put a perspective

458
00:30:44,360 --> 00:30:49,800
Speaker 1: around a problem statement. There are many regulatory documents such

459
00:30:49,800 --> 00:30:57,080
Speaker 1: as contracts or security filings which contains important clauses that

460
00:30:57,160 --> 00:31:02,120
Speaker 1: have really really serious business implications for example, payment terms,

461
00:31:02,680 --> 00:31:08,680
Speaker 1: obligations made to regulatory bodies, or warranties. Such humans can

462
00:31:08,720 --> 00:31:14,240
Speaker 1: spend countless hours reading and extracting the information so they

463
00:31:14,280 --> 00:31:18,560
Speaker 1: remain compliant. Although we can provide some of the out

464
00:31:18,560 --> 00:31:21,640
Speaker 1: of the box models for contracts and invoices and such,

465
00:31:22,120 --> 00:31:25,320
Speaker 1: but it creates UM but client may still need to

466
00:31:25,320 --> 00:31:28,920
Speaker 1: create their own element classifications of business classes. So the

467
00:31:29,080 --> 00:31:34,440
Speaker 1: solution has been with our debaters birth based classification technology

468
00:31:34,520 --> 00:31:37,720
Speaker 1: into these products so we can learn with few one

469
00:31:38,160 --> 00:31:42,960
Speaker 1: samples to do new classification of elements. Business documents could

470
00:31:43,000 --> 00:31:47,200
Speaker 1: include contracts, invoices, and procurement contracts. The end of the day,

471
00:31:47,240 --> 00:31:51,840
Speaker 1: it really really excelerates the outcomes what the businesses would

472
00:31:51,880 --> 00:31:58,720
Speaker 1: be looking for. UM. Other technology is around summarization. So

473
00:31:58,840 --> 00:32:02,040
Speaker 1: the problem statement here is like when you're looking for

474
00:32:02,720 --> 00:32:07,680
Speaker 1: information customer or employee who may have aggregate research from

475
00:32:07,720 --> 00:32:12,959
Speaker 1: different sources, clicking through multiple links and pages and finding

476
00:32:13,040 --> 00:32:17,000
Speaker 1: exactly what they need can be very very difficult, right.

477
00:32:17,000 --> 00:32:21,320
Speaker 1: It can take months, weeks and months sometimes to your years.

478
00:32:21,360 --> 00:32:25,880
Speaker 1: So with Watson and Debater technology, we can analyze variety

479
00:32:25,960 --> 00:32:29,600
Speaker 1: of these sources and provide a summary or brief of

480
00:32:29,680 --> 00:32:34,240
Speaker 1: the ideas and the information which is contained within UM

481
00:32:34,280 --> 00:32:37,120
Speaker 1: that's coming up. We're going to be leveraging this technology

482
00:32:37,120 --> 00:32:42,280
Speaker 1: in our Watson discovery portfolio in second half. The other

483
00:32:42,520 --> 00:32:47,160
Speaker 1: interesting UM issues we see today is like in our

484
00:32:47,200 --> 00:32:53,960
Speaker 1: traditional UH rule based systems for contact centers, it categorizes

485
00:32:54,280 --> 00:32:57,760
Speaker 1: large fraction of calls in a very generic bucket like

486
00:32:57,840 --> 00:33:01,160
Speaker 1: it says, you know, like not uncommon to see more

487
00:33:01,200 --> 00:33:05,400
Speaker 1: than maybe of calling a call center for a bank,

488
00:33:05,800 --> 00:33:08,160
Speaker 1: which says, hey, this this call was just made for

489
00:33:08,640 --> 00:33:12,480
Speaker 1: generalized checking, and it prevents the company from creating any

490
00:33:12,640 --> 00:33:17,280
Speaker 1: robust self service. So with Debater technology, now we can

491
00:33:17,360 --> 00:33:22,880
Speaker 1: leverage advanced topic clustering, which enables users to cluster this

492
00:33:23,040 --> 00:33:28,760
Speaker 1: incoming data in a meaningful topics of related information and

493
00:33:28,840 --> 00:33:33,680
Speaker 1: automatically this can be analyzed. So think of discovery of

494
00:33:33,760 --> 00:33:37,120
Speaker 1: a content minor which will be enhanced with this type

495
00:33:37,120 --> 00:33:42,080
Speaker 1: of a technology to extract better topics from very large

496
00:33:42,160 --> 00:33:46,640
Speaker 1: data sets and then make the topic extraction more business

497
00:33:46,720 --> 00:33:50,200
Speaker 1: user friendly. So a lot of stuff. I give a

498
00:33:50,240 --> 00:33:53,120
Speaker 1: lot of examples, but sort of the gist of all

499
00:33:53,200 --> 00:33:59,680
Speaker 1: this is, Look, it's going to impact businesses real outcomes, right,

500
00:33:59,760 --> 00:34:02,280
Speaker 1: It's going to save them time, is going to automate

501
00:34:02,320 --> 00:34:04,880
Speaker 1: the process, it's going to remove a lot of human

502
00:34:05,040 --> 00:34:09,880
Speaker 1: error which comes with it, and really speak towards the productivity.

503
00:34:10,200 --> 00:34:14,040
Speaker 1: Is going to speak towards the clients UM and P

504
00:34:14,200 --> 00:34:17,239
Speaker 1: as their own promoter scores and such, and so that's

505
00:34:17,280 --> 00:34:19,560
Speaker 1: really the gist of what we're looking to drive out

506
00:34:19,560 --> 00:34:22,920
Speaker 1: of the debater technology. If I'm understanding this correctly, this

507
00:34:23,000 --> 00:34:26,200
Speaker 1: is interesting that it's interesting that this kind of functionality

508
00:34:26,280 --> 00:34:30,000
Speaker 1: would come out of an AI debate tool, because debate

509
00:34:30,040 --> 00:34:32,600
Speaker 1: and persuasion that will seem like the kinds of things

510
00:34:32,640 --> 00:34:35,880
Speaker 1: that would be inherently the most difficult to master with AI,

511
00:34:35,960 --> 00:34:38,719
Speaker 1: because you've got all these elements of style and subtlety

512
00:34:39,120 --> 00:34:41,959
Speaker 1: things that are really difficult to quantify to make into

513
00:34:42,160 --> 00:34:45,960
Speaker 1: two understandable data. But out of the debater technology, it

514
00:34:46,000 --> 00:34:48,000
Speaker 1: sounds like you're saying that you're actually getting a lot

515
00:34:48,000 --> 00:34:52,080
Speaker 1: of derivative technologies that are good at dealing with algorithmic

516
00:34:52,200 --> 00:34:55,680
Speaker 1: types of text like legal documents. Am I getting this right?

517
00:34:55,800 --> 00:34:59,200
Speaker 1: Like that you could have a piece of software that

518
00:34:59,280 --> 00:35:02,920
Speaker 1: works like a lawyer. Uh, and it can explain this

519
00:35:03,080 --> 00:35:05,600
Speaker 1: contract to you when it's going over your head. And

520
00:35:05,840 --> 00:35:08,759
Speaker 1: this kind of thing is possible now because of how

521
00:35:08,840 --> 00:35:12,719
Speaker 1: formulaic and algorithmic legal documents tend to be. Would that

522
00:35:12,760 --> 00:35:16,040
Speaker 1: be a correct understanding? Yeah, no, totally And if I may,

523
00:35:16,360 --> 00:35:20,560
Speaker 1: UM give you one of the client example, especially as

524
00:35:20,560 --> 00:35:25,200
Speaker 1: you started talking about legal UM, Legal Nations platform actually

525
00:35:25,200 --> 00:35:29,719
Speaker 1: provides this in house legal teams and outside console the

526
00:35:29,719 --> 00:35:33,960
Speaker 1: ability to respond to their lawsuits UM and draft their

527
00:35:34,000 --> 00:35:37,720
Speaker 1: initial round up discovery requests literally less than two minutes

528
00:35:37,840 --> 00:35:43,279
Speaker 1: right um and which shaved off about ten hours of

529
00:35:43,480 --> 00:35:47,319
Speaker 1: attorney times on each of these lawsuits. So the real

530
00:35:47,400 --> 00:35:50,879
Speaker 1: direct outcomes of usage of this technology. So you've been

531
00:35:50,880 --> 00:35:53,919
Speaker 1: talking about big business applications, but I also wonder about

532
00:35:53,920 --> 00:35:58,160
Speaker 1: applications directly for the consumer. Where, for example, because you

533
00:35:58,200 --> 00:36:01,960
Speaker 1: have a program that ingests legal documents, so you you

534
00:36:02,040 --> 00:36:04,920
Speaker 1: feed it some contract you're thinking about signing, and then

535
00:36:04,960 --> 00:36:07,120
Speaker 1: you say, I have a question because I'm not a lawyer,

536
00:36:07,239 --> 00:36:09,759
Speaker 1: I don't understand what I would be bound to do

537
00:36:09,960 --> 00:36:13,080
Speaker 1: under this agreement. And then you could feed the contract

538
00:36:13,160 --> 00:36:16,960
Speaker 1: in and pose questions to your AI legal assistant in

539
00:36:17,080 --> 00:36:21,359
Speaker 1: natural language. Can you see a future like that. We do,

540
00:36:21,640 --> 00:36:25,319
Speaker 1: and we already have a product like what'son Assistant, which

541
00:36:25,320 --> 00:36:27,680
Speaker 1: is for customer care. It feeds on a lot of

542
00:36:27,760 --> 00:36:31,120
Speaker 1: you know, pre train models, like especially now in COVID

543
00:36:31,239 --> 00:36:37,720
Speaker 1: nineteen right, Uh, a situation where our government offices and

544
00:36:38,400 --> 00:36:42,000
Speaker 1: our healthcare are getting in dated by calls. Right, So

545
00:36:43,080 --> 00:36:48,000
Speaker 1: leveraging this UM what'son Assistant in front is really helping

546
00:36:48,040 --> 00:36:50,960
Speaker 1: them deflect a lot of those phone calls and get

547
00:36:51,000 --> 00:36:54,759
Speaker 1: the accurate answers in hands of the consumers. So you know,

548
00:36:55,120 --> 00:36:57,399
Speaker 1: this is what we are focusing on around customer care.

549
00:36:57,680 --> 00:37:00,640
Speaker 1: But yeah, in the future, I mean this similar technology

550
00:37:00,640 --> 00:37:05,960
Speaker 1: and leveraging UM the from debater, we can actually go

551
00:37:06,120 --> 00:37:10,440
Speaker 1: into any domain. Right, we have the right framework and

552
00:37:10,480 --> 00:37:15,000
Speaker 1: we have the right technology to go pursue those different domains.

553
00:37:15,040 --> 00:37:17,040
Speaker 1: I guess this sets us up for a bigger question,

554
00:37:17,160 --> 00:37:20,200
Speaker 1: which is what is the overall role of natural language

555
00:37:20,200 --> 00:37:24,040
Speaker 1: processing in the landscape of AI today and also which

556
00:37:24,040 --> 00:37:27,240
Speaker 1: are the elements of natural language processing that we've really

557
00:37:27,280 --> 00:37:29,239
Speaker 1: gotten a lot better at and which are the ones

558
00:37:29,280 --> 00:37:32,879
Speaker 1: that are still a major challenge. Yeah, great question. As

559
00:37:32,880 --> 00:37:35,719
Speaker 1: we all know, right, language have existed. I don't know

560
00:37:35,760 --> 00:37:39,720
Speaker 1: a hundred thousand plus years. You know, started as speech

561
00:37:39,800 --> 00:37:43,480
Speaker 1: probably people started to talk and the writing came perhaps

562
00:37:43,560 --> 00:37:47,359
Speaker 1: much later. Um, and we write in ways we don't

563
00:37:47,400 --> 00:37:52,000
Speaker 1: talk also, right, it's a lot more descriptive and more reflective.

564
00:37:52,760 --> 00:37:57,719
Speaker 1: And so now with things where we can compute at

565
00:37:57,840 --> 00:38:02,399
Speaker 1: larger with open data sets and transfer learnings, n LP

566
00:38:02,680 --> 00:38:06,719
Speaker 1: natural language processing really really is the inflection point, right,

567
00:38:06,760 --> 00:38:10,080
Speaker 1: And some of the examples I shared earlier around the

568
00:38:10,160 --> 00:38:16,720
Speaker 1: sentiment analysis and summarization and clustering, these are all such

569
00:38:16,880 --> 00:38:23,000
Speaker 1: critical aspects of taking LP, not just natural language processing,

570
00:38:23,040 --> 00:38:27,080
Speaker 1: but natural language understanding, natural language generations is all going

571
00:38:27,120 --> 00:38:30,160
Speaker 1: to come through all of that. And we really think

572
00:38:30,719 --> 00:38:34,440
Speaker 1: with the Debater technology it really puts us in a

573
00:38:35,160 --> 00:38:37,880
Speaker 1: in a leader quadrant here a lot more work to

574
00:38:37,920 --> 00:38:41,400
Speaker 1: be done, but the the end goal is yes, we

575
00:38:41,440 --> 00:38:44,280
Speaker 1: can continue to research on these things, but how quickly

576
00:38:44,360 --> 00:38:47,600
Speaker 1: we commercialize it and how quick quickly we help our

577
00:38:47,640 --> 00:38:51,200
Speaker 1: clients and users to see the outcomes what are needed

578
00:38:51,239 --> 00:38:54,520
Speaker 1: here and make them a lot more productive. So how

579
00:38:54,560 --> 00:38:58,879
Speaker 1: many languages does Project Debater and Watson together, how many

580
00:38:58,920 --> 00:39:03,400
Speaker 1: do they understand support today? We started with obviously English,

581
00:39:03,560 --> 00:39:10,200
Speaker 1: we are expanding now to French, Spanish, German in in

582
00:39:10,320 --> 00:39:12,880
Speaker 1: the second half of this year, and then very soon

583
00:39:12,960 --> 00:39:18,799
Speaker 1: will expand to Dutch, French, Arabic, Chinese both traditional and simplified,

584
00:39:19,360 --> 00:39:24,440
Speaker 1: and Italian. UM And obviously we are choosing these based

585
00:39:24,440 --> 00:39:26,920
Speaker 1: on where we are seeing most of our growth and

586
00:39:27,960 --> 00:39:31,560
Speaker 1: an adoption. What are some additional examples of how these

587
00:39:31,560 --> 00:39:35,839
Speaker 1: commercialized capabilities can be used by clients? Great question, um.

588
00:39:35,880 --> 00:39:40,000
Speaker 1: I gave you an example earlier on legal missions. The

589
00:39:40,080 --> 00:39:42,880
Speaker 1: other one, which is very close to my heart is

590
00:39:43,800 --> 00:39:50,640
Speaker 1: um RBS with Watson. Watson RBS built Cora, which is it?

591
00:39:50,800 --> 00:39:53,719
Speaker 1: Which is their digital assistant that helps better serve their

592
00:39:53,760 --> 00:39:59,719
Speaker 1: customers through first time problem resolutions. Cora is trained with

593
00:39:59,800 --> 00:40:06,200
Speaker 1: the were one thousand responses to more than two customer queries. However,

594
00:40:06,480 --> 00:40:09,160
Speaker 1: if she doesn't know an answer or she sends that

595
00:40:09,239 --> 00:40:13,759
Speaker 1: customer is getting angry or frustrated, she will transfer it

596
00:40:13,800 --> 00:40:18,680
Speaker 1: to a live agent. Now, with improved sentiment analysis from Debater,

597
00:40:18,840 --> 00:40:22,680
Speaker 1: as I mentioned earlier, we hope that clients like RBS

598
00:40:22,760 --> 00:40:26,920
Speaker 1: will be able to better serve their customers by having

599
00:40:27,080 --> 00:40:32,200
Speaker 1: digital assistance that better understand the subtleties of the of

600
00:40:32,280 --> 00:40:36,000
Speaker 1: the sentiments of the clients. So for example, the phrase

601
00:40:36,760 --> 00:40:41,080
Speaker 1: over the moon might be interpreted as literally about the

602
00:40:41,080 --> 00:40:47,000
Speaker 1: planetary satellite and not as excited or elated. Right. So

603
00:40:47,200 --> 00:40:50,600
Speaker 1: this is what with Project Debater core AI built into

604
00:40:50,680 --> 00:40:55,520
Speaker 1: IBM Watson, it can understand these idioms helping clients like

605
00:40:55,760 --> 00:41:00,560
Speaker 1: RBS to better serve their customers. The other example switching

606
00:41:00,560 --> 00:41:05,400
Speaker 1: into financial like Credit Mutual, they had over five thousand

607
00:41:05,480 --> 00:41:11,560
Speaker 1: branches and they receive more than three fifty thousand online

608
00:41:11,560 --> 00:41:15,399
Speaker 1: inquiries a day and the volume is growing at least

609
00:41:15,440 --> 00:41:19,080
Speaker 1: twenty three percent a year. So now with Watson infused

610
00:41:19,200 --> 00:41:25,360
Speaker 1: email analyzer, they can help deflect and address of the

611
00:41:25,440 --> 00:41:31,760
Speaker 1: three daily emails received by banks client advisors. So the

612
00:41:31,800 --> 00:41:36,480
Speaker 1: implementation of the topic clustering from Debater, we believe now

613
00:41:36,560 --> 00:41:40,600
Speaker 1: clients with similar needs that Credit Mutual will enable more

614
00:41:40,640 --> 00:41:45,600
Speaker 1: self service by identifying clusters are commonly as topics and

615
00:41:45,680 --> 00:41:50,200
Speaker 1: can be converted into self service content. Right. So to me,

616
00:41:50,320 --> 00:41:53,600
Speaker 1: the examples like this are just amazing because I can

617
00:41:53,680 --> 00:41:58,040
Speaker 1: totally then connect the dots between technology, the usage and

618
00:41:58,080 --> 00:42:02,360
Speaker 1: the outcome, right, a win win situation. We've got multiple

619
00:42:02,400 --> 00:42:06,040
Speaker 1: other examples as well, Roberts, and we're going to continue

620
00:42:06,040 --> 00:42:10,279
Speaker 1: to be focusing on how do we really not just

621
00:42:10,400 --> 00:42:15,480
Speaker 1: commercialize it, but I believe in AI is really meant

622
00:42:15,520 --> 00:42:19,759
Speaker 1: to improve our society as well, right, make us more

623
00:42:19,800 --> 00:42:23,279
Speaker 1: productive and do better things, especially the world we are

624
00:42:23,320 --> 00:42:25,920
Speaker 1: living in with COVID and other things which are happening

625
00:42:26,000 --> 00:42:29,520
Speaker 1: around us. Right, Um, the goodness of AI needs to

626
00:42:29,560 --> 00:42:33,640
Speaker 1: be there, so very critical overall, what do you see

627
00:42:33,680 --> 00:42:36,839
Speaker 1: as the best possible role for AI, not just as

628
00:42:36,840 --> 00:42:40,080
Speaker 1: a tool for business, but as a society. What could

629
00:42:40,160 --> 00:42:43,080
Speaker 1: it do for us in the best case scenario? Yeah,

630
00:42:43,160 --> 00:42:47,200
Speaker 1: I mean that's a great question, right um. To me fundamentally,

631
00:42:47,239 --> 00:42:50,280
Speaker 1: I mean there are many examples, but one most critical

632
00:42:50,360 --> 00:42:53,600
Speaker 1: which comes to my mind is how AI can really

633
00:42:53,640 --> 00:42:56,919
Speaker 1: help us detect bias? Right, A lot of our data

634
00:42:56,920 --> 00:43:03,520
Speaker 1: sets and it has been built by humans with unbiased

635
00:43:04,120 --> 00:43:08,000
Speaker 1: goes into those data. Right, AI can really start separating

636
00:43:08,040 --> 00:43:13,160
Speaker 1: that help us detect bias and and make our products better,

637
00:43:13,560 --> 00:43:16,360
Speaker 1: makes our society better. So that to me is the

638
00:43:17,480 --> 00:43:19,399
Speaker 1: would be sort of the holy grail if I can

639
00:43:19,440 --> 00:43:24,720
Speaker 1: achieve that. All right, So there you have it. Thanks

640
00:43:24,800 --> 00:43:27,960
Speaker 1: once again to know I'm slow name and Maduka char

641
00:43:28,120 --> 00:43:30,000
Speaker 1: for taking time out of their busy days to chat

642
00:43:30,040 --> 00:43:33,880
Speaker 1: with us about this topic. For more information on smart Talks,

643
00:43:34,040 --> 00:43:37,680
Speaker 1: visit IBM dot com slash smart Talks, and if you'd

644
00:43:37,680 --> 00:43:39,480
Speaker 1: like to learn more about n LP, you can go

645
00:43:39,520 --> 00:43:44,320
Speaker 1: to IBM dot com slash Watson, Slash Natural dash Language,

646
00:43:44,400 --> 00:43:47,239
Speaker 1: dash Processing. And if you would like to learn more

647
00:43:47,280 --> 00:43:49,799
Speaker 1: about our show, well, you can find us wherever you

648
00:43:49,840 --> 00:43:53,239
Speaker 1: get your podcasts and wherever that happens to be. Just

649
00:43:53,320 --> 00:43:58,040
Speaker 1: make sure you rate, review and subscribe. Huge thanks as

650
00:43:58,040 --> 00:44:01,480
Speaker 1: always to our excellent audio producers eth Nicholas Johnson. If

651
00:44:01,520 --> 00:44:02,960
Speaker 1: you would like to get in touch with us with

652
00:44:03,120 --> 00:44:05,680
Speaker 1: feedback on this episode or any other, to suggest a

653
00:44:05,719 --> 00:44:07,960
Speaker 1: topic for the future, or just to say hello, you

654
00:44:07,960 --> 00:44:10,799
Speaker 1: can email us at contact at stuff to Blow your

655
00:44:10,800 --> 00:44:21,120
Speaker 1: Mind dot com. Stuff to Blow Your Mind is production

656
00:44:21,160 --> 00:44:23,920
Speaker 1: of I Heart Radio. For more podcasts for my Heart Radio,

657
00:44:24,120 --> 00:44:26,800
Speaker 1: visit the i Heart Radio app, Apple Podcasts, or wherever

658
00:44:26,840 --> 00:44:36,120
Speaker 1: you listening to your favorite shows,