1 00:00:04,400 --> 00:00:12,600 Speaker 1: Welcome to tex Stuff production from I Heart Radio. Hey there, 2 00:00:12,640 --> 00:00:16,200 Speaker 1: and welcome to tech Stuff. I'm your host, Jonathan Strickland. 3 00:00:16,239 --> 00:00:18,400 Speaker 1: I'm an executive producer with I Heart Radio and a 4 00:00:18,480 --> 00:00:21,520 Speaker 1: love of all things tech, and I'm going to bring 5 00:00:21,560 --> 00:00:24,800 Speaker 1: you guys a little bit of a rerun today. I 6 00:00:24,840 --> 00:00:27,960 Speaker 1: am trying to get some stuff put together for a 7 00:00:28,040 --> 00:00:32,199 Speaker 1: special series of episodes as well as prepare for some 8 00:00:32,280 --> 00:00:35,199 Speaker 1: other stuff. So great things right around the corner. I 9 00:00:35,200 --> 00:00:37,600 Speaker 1: did not want to leave you without an episode at all, 10 00:00:37,680 --> 00:00:40,000 Speaker 1: So we're going to listen to this one that originally 11 00:00:40,040 --> 00:00:43,280 Speaker 1: published on October eighteenth, two thousand eighteen, and it kind 12 00:00:43,280 --> 00:00:45,559 Speaker 1: of goes in line with some other stuff we've been 13 00:00:45,560 --> 00:00:49,040 Speaker 1: covering in recent episodes of Tech Stuff. This episode was 14 00:00:49,080 --> 00:00:54,600 Speaker 1: titled Could We Make a Sarcastic Supercomputer? And yeah, it 15 00:00:54,640 --> 00:00:59,920 Speaker 1: really dives into the whole concept of artificial intelligence, natural lane, 16 00:01:00,040 --> 00:01:03,960 Speaker 1: which and just kind of understanding the quirks of what 17 00:01:04,040 --> 00:01:07,120 Speaker 1: it is to be human and the whole concept of sarcasm. 18 00:01:07,360 --> 00:01:10,680 Speaker 1: I hope you guys enjoy it. I mean that without 19 00:01:10,720 --> 00:01:13,280 Speaker 1: even a hint of sarcasm. And I'll chat with you 20 00:01:13,480 --> 00:01:18,200 Speaker 1: after the episode. Today. I want to talk to you 21 00:01:18,319 --> 00:01:21,920 Speaker 1: about an interesting topic that I got to explore a 22 00:01:21,959 --> 00:01:25,760 Speaker 1: couple of years ago with Joe McCormick and Lauren fogobaum 23 00:01:26,200 --> 00:01:30,880 Speaker 1: As we debated the possibilities of computers learning how to 24 00:01:31,280 --> 00:01:36,440 Speaker 1: understand sarcasm. We did it for a podcast called Forward Thinking, 25 00:01:36,760 --> 00:01:38,760 Speaker 1: which was around for a couple of years. It was 26 00:01:38,800 --> 00:01:40,760 Speaker 1: a lot of fun to work on that that show 27 00:01:41,319 --> 00:01:43,840 Speaker 1: is over, but I thought I would revisit the topic 28 00:01:44,200 --> 00:01:47,160 Speaker 1: and talk about it for you guys and kind of 29 00:01:47,200 --> 00:01:50,360 Speaker 1: go over what would it take to have a computer 30 00:01:50,440 --> 00:01:54,360 Speaker 1: that could actually understand when someone's being sarcastic. Now to 31 00:01:54,840 --> 00:01:57,360 Speaker 1: understand why this is a big deal, it helps to 32 00:01:57,440 --> 00:02:01,520 Speaker 1: have a refresher course on how computers process information. And 33 00:02:01,560 --> 00:02:03,960 Speaker 1: I know I talked about this a lot, but I 34 00:02:04,000 --> 00:02:07,000 Speaker 1: still think it's important to cover the basics when you 35 00:02:07,000 --> 00:02:10,160 Speaker 1: want to talk about something as advanced as being able 36 00:02:10,200 --> 00:02:16,200 Speaker 1: to detect and understand sarcasm. So computers understand machine code 37 00:02:16,280 --> 00:02:19,960 Speaker 1: or assembly language. This is a language that corresponds with 38 00:02:20,080 --> 00:02:25,400 Speaker 1: the actual physical architecture of the computers. So the way 39 00:02:25,440 --> 00:02:28,639 Speaker 1: the computer is built, that's how this language interacts. It's 40 00:02:28,680 --> 00:02:32,440 Speaker 1: it's essentially how the physical components of the computer are 41 00:02:32,520 --> 00:02:38,480 Speaker 1: able to handle electric current or voltage differences in order 42 00:02:38,520 --> 00:02:45,600 Speaker 1: to process information, and computers can interpret this and execute 43 00:02:45,720 --> 00:02:49,919 Speaker 1: upon this language very quickly. It is the basic language 44 00:02:49,960 --> 00:02:55,600 Speaker 1: of those physical components. However, it is almost impossible for 45 00:02:55,960 --> 00:02:58,320 Speaker 1: humans to work with this, at least on a way 46 00:02:58,400 --> 00:03:02,720 Speaker 1: that is at all of shion, because it ultimately for 47 00:03:02,919 --> 00:03:08,520 Speaker 1: most computers boils down to binary language, right, zeros and ones. 48 00:03:09,360 --> 00:03:13,079 Speaker 1: So you see a huge block of zeros and ones, 49 00:03:13,120 --> 00:03:15,640 Speaker 1: and unless you are neo from the matrix, it means 50 00:03:15,680 --> 00:03:20,080 Speaker 1: nothing to you. So we speak in natural language to 51 00:03:20,120 --> 00:03:24,120 Speaker 1: one another. Natural language, however, is filled with a lot 52 00:03:24,160 --> 00:03:28,600 Speaker 1: of components that make it very very challenging for machines 53 00:03:28,639 --> 00:03:33,280 Speaker 1: to interpret, like ambiguity, or there might be double meanings 54 00:03:33,320 --> 00:03:36,320 Speaker 1: in a phrase and you may mean both meanings at 55 00:03:36,320 --> 00:03:40,640 Speaker 1: the same time, and that is too complicated for most 56 00:03:40,680 --> 00:03:43,680 Speaker 1: machines to be able to process. They just can't deal 57 00:03:43,760 --> 00:03:47,520 Speaker 1: with that. So to bridge the gap between the way 58 00:03:47,600 --> 00:03:51,840 Speaker 1: we humans communicate and the way that computers process language, 59 00:03:52,280 --> 00:03:58,000 Speaker 1: we have created programming languages and compilers. Now, programming languages 60 00:03:58,080 --> 00:04:02,120 Speaker 1: fall into two broad category worries. It's more like a spectrum, 61 00:04:02,440 --> 00:04:04,840 Speaker 1: and you could be further on one end than the other, 62 00:04:05,280 --> 00:04:08,920 Speaker 1: and we typically call them high level programming languages and 63 00:04:09,000 --> 00:04:13,920 Speaker 1: low level programming languages. The lower the level of programming language, 64 00:04:14,000 --> 00:04:17,719 Speaker 1: the closer it is to machine code, and the easier 65 00:04:17,760 --> 00:04:20,800 Speaker 1: it is for a computer to understand, but the harder 66 00:04:20,839 --> 00:04:22,880 Speaker 1: it is to work with. If you happen to be, 67 00:04:22,960 --> 00:04:27,240 Speaker 1: you know, a human being. High level programming languages are 68 00:04:27,320 --> 00:04:30,640 Speaker 1: easier for humans to understand. Now, if you have never 69 00:04:30,720 --> 00:04:33,919 Speaker 1: taken any courses in programming and you're looking at a 70 00:04:34,000 --> 00:04:38,039 Speaker 1: page of code, it could seem indecipherable to you. It 71 00:04:38,200 --> 00:04:44,160 Speaker 1: is just meaningless strings of characters. But once you learn 72 00:04:44,200 --> 00:04:49,200 Speaker 1: the rules of that programming language, how you construct an instruction, 73 00:04:49,640 --> 00:04:51,720 Speaker 1: and a series of instructions, how you go from one 74 00:04:51,720 --> 00:04:55,080 Speaker 1: instruction to the next. Once you understand the rules, it 75 00:04:55,120 --> 00:04:58,159 Speaker 1: actually becomes quite easy to use in the grand scheme 76 00:04:58,160 --> 00:05:00,640 Speaker 1: of things, much more easy than machine which would be. 77 00:05:01,720 --> 00:05:04,680 Speaker 1: But again, the problem here is that computers don't understand 78 00:05:04,720 --> 00:05:09,880 Speaker 1: programming languages, not natively. Even though this is not exactly 79 00:05:09,920 --> 00:05:12,320 Speaker 1: the same as human natural language, it's also not the 80 00:05:12,400 --> 00:05:15,920 Speaker 1: same as machine language. That's why you need compilers. A 81 00:05:15,920 --> 00:05:21,840 Speaker 1: compiler is essentially a translator. It takes this high level 82 00:05:21,920 --> 00:05:26,240 Speaker 1: programming language or higher level anyway and then converts it 83 00:05:26,320 --> 00:05:29,440 Speaker 1: into a machine readable language for the computer to actually 84 00:05:29,440 --> 00:05:32,680 Speaker 1: execute upon. And this is all in the design of 85 00:05:32,680 --> 00:05:37,320 Speaker 1: the programming languages and the compilers. So this is the 86 00:05:37,400 --> 00:05:41,960 Speaker 1: way that for decades we have interacted with computers, when 87 00:05:41,960 --> 00:05:43,960 Speaker 1: you're talking about it on a on a direct level, 88 00:05:44,000 --> 00:05:48,080 Speaker 1: not just executing a program, but creating code, creating programs 89 00:05:48,080 --> 00:05:53,160 Speaker 1: for computers to run. Over the last few decades, we've 90 00:05:53,200 --> 00:05:58,880 Speaker 1: had some very very smart people working on natural language 91 00:05:58,880 --> 00:06:05,760 Speaker 1: systems for machines, which would allow a computer to interpret 92 00:06:06,360 --> 00:06:09,880 Speaker 1: natural language in a way that would make some sort 93 00:06:09,920 --> 00:06:11,640 Speaker 1: of sense, and for the computer to be able to 94 00:06:11,680 --> 00:06:15,280 Speaker 1: act upon that language. And we've seen this in plenty 95 00:06:15,320 --> 00:06:20,640 Speaker 1: of examples recently. Most smartphones have some sort of smart assistant. 96 00:06:21,120 --> 00:06:25,279 Speaker 1: You have standalone products like Amazon's Echo, you have Google Home, 97 00:06:25,400 --> 00:06:30,560 Speaker 1: You've got tons of devices that can interact with people. 98 00:06:30,839 --> 00:06:35,600 Speaker 1: It can be activated by typically an alert phrase, which 99 00:06:35,600 --> 00:06:37,279 Speaker 1: I'm not going to say because I don't want any 100 00:06:37,320 --> 00:06:39,120 Speaker 1: of you guys to have to deal with that. I 101 00:06:39,200 --> 00:06:41,559 Speaker 1: know how irritating it is when I'm watching a video 102 00:06:41,680 --> 00:06:48,120 Speaker 1: and someone activates their specific system and then mine begins 103 00:06:48,160 --> 00:06:50,200 Speaker 1: to respond, and all my lights started going on and 104 00:06:50,279 --> 00:06:53,760 Speaker 1: off because the people on YouTube we're talking funny. I 105 00:06:53,800 --> 00:06:56,640 Speaker 1: know how irritating that is, but use that it activates 106 00:06:56,920 --> 00:07:00,839 Speaker 1: and then you can speak. And typically you can say 107 00:07:00,880 --> 00:07:06,279 Speaker 1: the same thing several different ways and the device appears 108 00:07:06,279 --> 00:07:09,240 Speaker 1: to understand you no matter how you word it. And 109 00:07:09,320 --> 00:07:11,960 Speaker 1: this is a real challenge because we human beings can 110 00:07:12,000 --> 00:07:14,760 Speaker 1: find lots of different ways to say the same thing. 111 00:07:15,120 --> 00:07:17,800 Speaker 1: For example, if I say what is the weather today, 112 00:07:18,640 --> 00:07:20,640 Speaker 1: it could be very similar to if I if I 113 00:07:20,680 --> 00:07:24,040 Speaker 1: ask a question is it going to rain today? Both 114 00:07:24,080 --> 00:07:27,280 Speaker 1: of those are asking for information about the weather, but 115 00:07:27,400 --> 00:07:31,160 Speaker 1: are very different ways of saying that. A good natural 116 00:07:31,280 --> 00:07:34,560 Speaker 1: language recognition program will be able to parse that information 117 00:07:35,320 --> 00:07:40,760 Speaker 1: and then return the appropriate response. This is not an 118 00:07:40,760 --> 00:07:45,040 Speaker 1: easy thing to do. Typically it involves creating a neural 119 00:07:45,080 --> 00:07:49,000 Speaker 1: network structure, and I've talked about artificial neural networks recently. 120 00:07:49,640 --> 00:07:56,520 Speaker 1: That's a typically a network that can accept multiple binary inputs, 121 00:07:56,560 --> 00:07:59,600 Speaker 1: so either a zero or a one input that represents 122 00:07:59,680 --> 00:08:03,760 Speaker 1: some thing uh, some sort of yes, no or on 123 00:08:03,920 --> 00:08:09,000 Speaker 1: off kind of feature. It can accept multiple multiple inputs 124 00:08:09,160 --> 00:08:12,360 Speaker 1: of that nature, so multiple zeros or ones that all 125 00:08:12,600 --> 00:08:15,520 Speaker 1: factor into making a decision, and then it has a 126 00:08:15,560 --> 00:08:19,800 Speaker 1: waiting for each of those components, and then it produces 127 00:08:19,880 --> 00:08:22,960 Speaker 1: a single output that's also binary in nature, either is 128 00:08:23,040 --> 00:08:26,320 Speaker 1: zero one, and it passes that on to other artificial 129 00:08:26,360 --> 00:08:30,040 Speaker 1: neurons further down the chain. Sometimes that will come back 130 00:08:30,080 --> 00:08:33,760 Speaker 1: around and you have a recursive artificial neural network. The 131 00:08:33,880 --> 00:08:40,080 Speaker 1: goal here is for this process two ultimately result in 132 00:08:40,640 --> 00:08:46,000 Speaker 1: a response that is reasonably certain to meet the requirements 133 00:08:46,040 --> 00:08:49,199 Speaker 1: of the person asking the question. This tends to be 134 00:08:49,720 --> 00:08:53,240 Speaker 1: talked about in the realm of probabilities. We we talked 135 00:08:53,240 --> 00:08:56,680 Speaker 1: about how certain the machine is that the response is 136 00:08:56,720 --> 00:09:00,440 Speaker 1: the appropriate one, and if it falls below a certain toushold, 137 00:09:01,120 --> 00:09:04,079 Speaker 1: then the machine would typically respond with I'm sorry, I 138 00:09:04,160 --> 00:09:06,600 Speaker 1: don't know what you're asking for, or something similar to that. 139 00:09:08,000 --> 00:09:10,520 Speaker 1: There are cases where you just get misinterpreted and you'll 140 00:09:10,520 --> 00:09:13,240 Speaker 1: get a response that does not reflect whatever you ask. 141 00:09:13,400 --> 00:09:16,080 Speaker 1: That's a little different. That's where the machine has drawn 142 00:09:16,120 --> 00:09:19,320 Speaker 1: a conclusion, has been reasonably certain that it came to 143 00:09:19,320 --> 00:09:21,320 Speaker 1: the right conclusion. It turns out it was wrong the 144 00:09:21,320 --> 00:09:26,000 Speaker 1: whole way, but that's the process. Now, when it comes 145 00:09:26,040 --> 00:09:34,000 Speaker 1: to sarcasm, that adds yet another layer of difficulty, because 146 00:09:34,400 --> 00:09:38,360 Speaker 1: now a machine isn't just parsing what you are saying. 147 00:09:38,760 --> 00:09:42,720 Speaker 1: It has to understand what you mean, the meaning of 148 00:09:42,760 --> 00:09:47,679 Speaker 1: your words and the meaning of the way you deliver them. 149 00:09:47,720 --> 00:09:50,320 Speaker 1: It could be different. So if I were to just 150 00:09:50,480 --> 00:09:55,480 Speaker 1: write out a phrase with no tone, nobody language, uh, 151 00:09:55,840 --> 00:10:00,200 Speaker 1: not emphasizing any one word over another, it might be 152 00:10:00,320 --> 00:10:04,520 Speaker 1: very difficult to detect what my intent was. It may 153 00:10:04,559 --> 00:10:07,760 Speaker 1: seem like I'm being sincere, when in fact I'm being insincere. 154 00:10:08,080 --> 00:10:12,520 Speaker 1: For example, Uh, if I were to say that guy 155 00:10:12,600 --> 00:10:18,240 Speaker 1: is super tall, but I'm being sarcastic, then just in 156 00:10:18,280 --> 00:10:21,640 Speaker 1: that phrase the way I write it out, you would think, oh, well, 157 00:10:21,679 --> 00:10:26,160 Speaker 1: that person he's looking at must be super tall. How 158 00:10:26,200 --> 00:10:30,319 Speaker 1: do you recognize sarcasm? How can you detect that this 159 00:10:30,400 --> 00:10:33,520 Speaker 1: is in place and then understand what the meaning underneath 160 00:10:33,520 --> 00:10:37,960 Speaker 1: it is. One of the approaches that has been put 161 00:10:38,000 --> 00:10:44,720 Speaker 1: forward relates to IBM's Watson platform. Now. Watson first made 162 00:10:44,720 --> 00:10:48,680 Speaker 1: headlines back when it was a contestant on Jeopardy. It 163 00:10:48,960 --> 00:10:53,120 Speaker 1: went up against two former champions, including Ken Jennings, who 164 00:10:53,200 --> 00:10:56,440 Speaker 1: shows up on a house Stuff Works podcast. Anyway, Watson 165 00:10:56,480 --> 00:11:00,120 Speaker 1: went up against these two former champions, and it is 166 00:11:00,160 --> 00:11:03,400 Speaker 1: able to interpret natural language. It had to in order 167 00:11:03,400 --> 00:11:05,400 Speaker 1: to play the game of Jeopardy, And for those who 168 00:11:05,440 --> 00:11:08,120 Speaker 1: do not know what jeopardy is or they're not familiar 169 00:11:08,160 --> 00:11:11,319 Speaker 1: with the game show, Jeopardy is a game where you 170 00:11:11,360 --> 00:11:17,320 Speaker 1: are presented with categories of trivia and each category has 171 00:11:17,400 --> 00:11:23,920 Speaker 1: multiple uh questions or multiple entries in it, and they 172 00:11:24,040 --> 00:11:29,600 Speaker 1: range in dollar value, and the lower dollar value ones 173 00:11:29,640 --> 00:11:33,240 Speaker 1: are easier to answer than the higher dollar value ones, 174 00:11:34,320 --> 00:11:37,920 Speaker 1: and UH, you're Typically the way jeopardy works is that 175 00:11:37,960 --> 00:11:40,840 Speaker 1: you're you're given quote unquote the answer and you have 176 00:11:40,920 --> 00:11:46,199 Speaker 1: to provide the question. So, uh, if the answer were 177 00:11:47,600 --> 00:11:53,640 Speaker 1: this film that detailed the adventures of a young playwright 178 00:11:53,880 --> 00:11:57,560 Speaker 1: in sixteenth century England, one best picture, you would say, 179 00:11:57,880 --> 00:12:02,160 Speaker 1: what was Shakespeare in Love? So this computer is playing 180 00:12:02,160 --> 00:12:04,480 Speaker 1: against these two former champions. This was sort of an 181 00:12:04,520 --> 00:12:09,800 Speaker 1: exhibition series of games. It wasn't meant for, uh, a 182 00:12:09,840 --> 00:12:12,600 Speaker 1: competition in the way the typical Jeopardy games were there 183 00:12:12,640 --> 00:12:16,440 Speaker 1: was money on the line. Was an exhibition and Watson won. 184 00:12:16,880 --> 00:12:19,480 Speaker 1: It beat both of the champions, and it did what 185 00:12:19,559 --> 00:12:23,160 Speaker 1: I was telling you. It It would analyze the clue 186 00:12:23,600 --> 00:12:27,200 Speaker 1: that was given, the answer that was given, It would 187 00:12:27,240 --> 00:12:30,199 Speaker 1: try and generate a question to correspond with that answer, 188 00:12:30,559 --> 00:12:33,720 Speaker 1: and only if the question met a certain threshold of 189 00:12:33,760 --> 00:12:37,120 Speaker 1: confidence with Watson buzz in. If it did not meet 190 00:12:37,160 --> 00:12:41,280 Speaker 1: that level of confidence, Watson would remain quiet. And most importantly, 191 00:12:41,520 --> 00:12:44,160 Speaker 1: Watson was not at all connected to the Internet. All 192 00:12:44,240 --> 00:12:49,880 Speaker 1: the information was contained within a massive series of servers 193 00:12:50,760 --> 00:12:53,320 Speaker 1: more than gosh I can't even remember. There's a ton 194 00:12:53,400 --> 00:12:58,680 Speaker 1: of processors attached to it. Um so a very powerful machine, 195 00:12:59,720 --> 00:13:05,880 Speaker 1: but it still wasn't exactly able to detect sarcasm. It 196 00:13:05,960 --> 00:13:10,240 Speaker 1: could work with word play and it could work with riddles, 197 00:13:10,280 --> 00:13:13,200 Speaker 1: so that was really impressive. But what it really did 198 00:13:13,240 --> 00:13:15,800 Speaker 1: was it gave IBM the opportunity to say, we have 199 00:13:15,960 --> 00:13:20,600 Speaker 1: this platform here and we're welcoming developers to create applications 200 00:13:20,600 --> 00:13:24,360 Speaker 1: that tap into this platform and make use of this 201 00:13:25,120 --> 00:13:28,800 Speaker 1: in order to do interesting stuff with it. And IBM 202 00:13:28,960 --> 00:13:31,559 Speaker 1: was largely working with the medical industry at that point 203 00:13:31,600 --> 00:13:37,839 Speaker 1: to try and help doctors treat and diagnose patients, and 204 00:13:37,920 --> 00:13:40,000 Speaker 1: it was sort of computer guidance. It wasn't that you 205 00:13:40,040 --> 00:13:44,200 Speaker 1: had an automatic doctor, but rather the doctor had what 206 00:13:44,559 --> 00:13:49,720 Speaker 1: equates to a medical expert to confer with when trying 207 00:13:49,760 --> 00:13:53,000 Speaker 1: to determine why's the best course of action for a patient. 208 00:13:54,000 --> 00:13:58,000 Speaker 1: IBM put up an application Program Interface or API, and 209 00:13:58,080 --> 00:14:02,720 Speaker 1: let developers create their own cognitive computing applications built on 210 00:14:02,840 --> 00:14:08,000 Speaker 1: top of Watson. One of those was called the tone analyzer. 211 00:14:08,240 --> 00:14:11,439 Speaker 1: It still exists back when we were doing this episode 212 00:14:11,440 --> 00:14:15,679 Speaker 1: for Forward Thinking. It was in the form of analyzing 213 00:14:15,720 --> 00:14:18,720 Speaker 1: some text and telling you whether or not that text 214 00:14:18,720 --> 00:14:23,760 Speaker 1: would come across as agreeable or argumentative, or positive or negative, 215 00:14:24,240 --> 00:14:28,880 Speaker 1: and it would assign tone to those pieces. I'll explain 216 00:14:28,960 --> 00:14:32,160 Speaker 1: more about how it did and what it did in 217 00:14:32,240 --> 00:14:34,480 Speaker 1: just a minute, but first let's take a quick break 218 00:14:34,680 --> 00:14:46,040 Speaker 1: to thank our sponsor. So how did this tone analyzer work. 219 00:14:46,520 --> 00:14:52,960 Speaker 1: It would search for cues in any written text, social cues, 220 00:14:53,040 --> 00:14:56,960 Speaker 1: written cues, emotional cues in order to determine the overall 221 00:14:57,040 --> 00:14:59,960 Speaker 1: tone of a piece, which actually meant that the analy 222 00:15:00,040 --> 00:15:05,200 Speaker 1: lizer would tag individual words within a text, words that 223 00:15:05,280 --> 00:15:11,000 Speaker 1: it recognized and had already pre labeled as falling into 224 00:15:11,080 --> 00:15:14,560 Speaker 1: various categories. So words that might have a positive meaning 225 00:15:14,960 --> 00:15:20,600 Speaker 1: like happy, glad, joy, things like that, those would get 226 00:15:20,680 --> 00:15:24,600 Speaker 1: tagged as cheerful. But then it would then assign all 227 00:15:24,640 --> 00:15:28,560 Speaker 1: the individual words tags and then tally everything up. So 228 00:15:28,840 --> 00:15:31,000 Speaker 1: let's say you've got a bunch of sentences and it 229 00:15:31,080 --> 00:15:36,960 Speaker 1: starts individually labeling certain words as being cheerful or sad, 230 00:15:37,120 --> 00:15:41,000 Speaker 1: or angry or helpful, and then it adds it all 231 00:15:41,080 --> 00:15:43,520 Speaker 1: up and then would give you a percentage. So a 232 00:15:43,560 --> 00:15:49,520 Speaker 1: message might be agreeable or thirty conscientious, you would actually 233 00:15:49,560 --> 00:15:52,640 Speaker 1: get multiples of these, and that would just really indicate 234 00:15:52,680 --> 00:15:58,240 Speaker 1: the density of those types of words within the message itself. Now, 235 00:15:58,920 --> 00:16:02,000 Speaker 1: in an ideal world, if language were very simple to 236 00:16:02,800 --> 00:16:07,160 Speaker 1: understand and interpret by machines, this would help you gauge 237 00:16:07,360 --> 00:16:11,200 Speaker 1: how people would respond to your work. Right, So, you 238 00:16:11,200 --> 00:16:15,240 Speaker 1: could write a message. Before you send it, you put 239 00:16:15,280 --> 00:16:18,240 Speaker 1: it through the tone analyzer and it tells you what 240 00:16:18,440 --> 00:16:22,480 Speaker 1: sort of a tone you are setting. So if you 241 00:16:22,520 --> 00:16:25,680 Speaker 1: wanted to create a business letter, you could send it 242 00:16:25,680 --> 00:16:27,560 Speaker 1: through this tone analyzer and if it came back as 243 00:16:27,600 --> 00:16:32,200 Speaker 1: saying it's coming across as as a indecisive, you might 244 00:16:32,240 --> 00:16:35,360 Speaker 1: want to go back in and edit that message so 245 00:16:35,400 --> 00:16:40,200 Speaker 1: that you can make a more straightforward and uh decisive 246 00:16:40,400 --> 00:16:43,800 Speaker 1: message and not give the wrong impression before you send 247 00:16:43,880 --> 00:16:47,160 Speaker 1: the message out to your actual human recipient and come 248 00:16:47,240 --> 00:16:49,760 Speaker 1: up with alternate word choices in order to make sure 249 00:16:49,800 --> 00:16:52,000 Speaker 1: that your message is received the way you intended it 250 00:16:52,440 --> 00:16:55,240 Speaker 1: and anyone who has communicated over the Internet can think 251 00:16:55,320 --> 00:16:58,040 Speaker 1: of ways that this might have been helpful in the past, 252 00:16:58,160 --> 00:17:03,600 Speaker 1: because again, language depends on so many different elements to 253 00:17:03,760 --> 00:17:06,840 Speaker 1: get your meaning across, and when you reduce it to 254 00:17:07,920 --> 00:17:11,040 Speaker 1: the written form, especially the written form online, where we 255 00:17:11,119 --> 00:17:15,639 Speaker 1: tend to be very short with our our communication, it 256 00:17:15,840 --> 00:17:19,200 Speaker 1: comes in very quick bursts, a couple of sentences here 257 00:17:19,280 --> 00:17:22,400 Speaker 1: or there. We lack all that body language, we lack 258 00:17:22,480 --> 00:17:25,720 Speaker 1: that tone. It's very easy to misinterpret. I'm sure there's 259 00:17:25,760 --> 00:17:28,880 Speaker 1: been an example in your life where either you got 260 00:17:28,880 --> 00:17:31,520 Speaker 1: offended from receiving something that was meant in a way 261 00:17:31,760 --> 00:17:34,520 Speaker 1: that was different from the way you you interpreted it, 262 00:17:34,640 --> 00:17:37,080 Speaker 1: or the reverse happened where you sent a message and 263 00:17:37,160 --> 00:17:41,440 Speaker 1: somebody had a reaction you did not anticipate because they 264 00:17:41,520 --> 00:17:44,400 Speaker 1: could not tell what tone you were using just from 265 00:17:44,440 --> 00:17:48,119 Speaker 1: the words you were using. Machines have that same problem. 266 00:17:48,359 --> 00:17:51,879 Speaker 1: In the future, an analyzer like this tone analyzer, it 267 00:17:51,920 --> 00:17:56,560 Speaker 1: could be incorporated into word processors, or email servers, or 268 00:17:56,920 --> 00:18:00,520 Speaker 1: email services, i should say, or social media platform. So 269 00:18:00,560 --> 00:18:03,320 Speaker 1: you start typing in your message and before you hit 270 00:18:03,760 --> 00:18:07,320 Speaker 1: published or post or send, you could analyze that text. 271 00:18:07,840 --> 00:18:09,719 Speaker 1: It could tell you what the tone is and then 272 00:18:09,760 --> 00:18:12,600 Speaker 1: you could say, oh, no, that's gonna come across totally 273 00:18:12,760 --> 00:18:15,000 Speaker 1: the wrong way, and you could actually fix it before 274 00:18:15,080 --> 00:18:17,160 Speaker 1: you posted it or sent it, and then you wouldn't 275 00:18:17,200 --> 00:18:20,840 Speaker 1: have that awkward decision of whether or not to edit something, or, 276 00:18:20,840 --> 00:18:23,800 Speaker 1: in the case of Twitter, which continues to refuse to 277 00:18:23,840 --> 00:18:27,080 Speaker 1: allow you to edit tweets, to delete a tweet. I 278 00:18:27,160 --> 00:18:30,119 Speaker 1: deleted a tweet the other day when I posted a 279 00:18:30,200 --> 00:18:32,840 Speaker 1: link to a news story, and I had done a 280 00:18:32,920 --> 00:18:36,240 Speaker 1: rookie mistake, one that I try to avoid, but I 281 00:18:36,800 --> 00:18:39,960 Speaker 1: did it this pastime, which is that I didn't think 282 00:18:40,000 --> 00:18:42,199 Speaker 1: to look at the date when the news item had 283 00:18:42,240 --> 00:18:45,400 Speaker 1: been published, and had been published a full year earlier, 284 00:18:45,760 --> 00:18:48,080 Speaker 1: so it was not new news, it was old news. 285 00:18:48,600 --> 00:18:51,399 Speaker 1: And uh then deleted the tweet and it wasn't up 286 00:18:51,440 --> 00:18:53,680 Speaker 1: for long, but I still felt dumb about it. It 287 00:18:53,680 --> 00:18:55,399 Speaker 1: would have been nice to have been able to check that. 288 00:18:55,600 --> 00:18:58,280 Speaker 1: Although that's not tone obviously, that's but similar in the 289 00:18:59,000 --> 00:19:02,360 Speaker 1: and the idea that you want to check before you 290 00:19:03,080 --> 00:19:06,400 Speaker 1: end up offending someone, unless you're one of those jerk 291 00:19:06,480 --> 00:19:08,880 Speaker 1: faces that just sets out to offend people, in which 292 00:19:08,880 --> 00:19:12,760 Speaker 1: case rethink your strategy. There are better things to do. 293 00:19:12,920 --> 00:19:15,080 Speaker 1: It's just as you can make just as big an 294 00:19:15,080 --> 00:19:17,520 Speaker 1: impact being a positive person as you can being a 295 00:19:17,560 --> 00:19:20,000 Speaker 1: jerk face. I know it can seem like it's more work, 296 00:19:20,080 --> 00:19:22,359 Speaker 1: but it's also more rewarding in the long run. Okay, 297 00:19:22,400 --> 00:19:25,720 Speaker 1: soapbox done. So. There is a demo of the tone 298 00:19:25,760 --> 00:19:30,119 Speaker 1: analyzer that's available online, and back when we were recording 299 00:19:30,400 --> 00:19:33,840 Speaker 1: Forward Thinking, the demo worked in a way where it 300 00:19:33,840 --> 00:19:38,080 Speaker 1: would tell you about emotional tone and break it down 301 00:19:38,080 --> 00:19:41,120 Speaker 1: by percentage. It's a little different now, but I want 302 00:19:41,119 --> 00:19:45,000 Speaker 1: to tell you the what words and the results we 303 00:19:45,080 --> 00:19:48,680 Speaker 1: got in the past because they were so much fun. 304 00:19:49,280 --> 00:19:51,359 Speaker 1: Granted you would get a different result now because the 305 00:19:51,359 --> 00:19:55,399 Speaker 1: tone analyzer has been tweaked since we recorded that episode. So, 306 00:19:55,440 --> 00:19:58,760 Speaker 1: when we recorded that episode, one of my co hosts 307 00:19:59,119 --> 00:20:02,600 Speaker 1: decided to put a a sentence that is somewhat known 308 00:20:02,720 --> 00:20:06,000 Speaker 1: in literary circles into this tone analyzer and find out 309 00:20:06,000 --> 00:20:09,280 Speaker 1: what it said. And the sentence used was it is 310 00:20:09,320 --> 00:20:13,040 Speaker 1: a truth universally acknowledged that a single man in possession 311 00:20:13,040 --> 00:20:16,680 Speaker 1: of a good fortune must be in want of a wife. Now, 312 00:20:16,680 --> 00:20:21,879 Speaker 1: the analyzer said that this emotional tone was cheerful, the 313 00:20:21,960 --> 00:20:26,720 Speaker 1: social tone was seventy six percent open and fifty one agreeable, 314 00:20:26,960 --> 00:20:31,360 Speaker 1: and the writing tone was analytical. You can also view 315 00:20:31,400 --> 00:20:34,200 Speaker 1: the sentence in terms of word count as opposed to 316 00:20:34,240 --> 00:20:37,359 Speaker 1: the weighted value of individual words, and using that view, 317 00:20:37,440 --> 00:20:41,720 Speaker 1: five percent of the sentence sentences were in an emotional tone, 318 00:20:41,760 --> 00:20:44,399 Speaker 1: eighty nine percent in a social tone, and five percent 319 00:20:44,440 --> 00:20:48,159 Speaker 1: in a writing tone. Now, the analyzer highlights each word 320 00:20:48,200 --> 00:20:52,600 Speaker 1: according to how it classifies them, so emotional words would 321 00:20:52,600 --> 00:20:55,879 Speaker 1: be highlighted in red or pink in that older version 322 00:20:55,880 --> 00:20:58,919 Speaker 1: of the Tone Analyzer, social words would show up in blue, 323 00:20:59,359 --> 00:21:02,960 Speaker 1: and writing tones would be in green. And you could 324 00:21:02,960 --> 00:21:05,800 Speaker 1: click on any word and the analyzer would offer alternative 325 00:21:05,880 --> 00:21:08,320 Speaker 1: words that you might want to use and classify those 326 00:21:08,320 --> 00:21:11,520 Speaker 1: words in the tones that they are associated with, so 327 00:21:11,560 --> 00:21:13,520 Speaker 1: that you could shape your message to meet the tone 328 00:21:13,520 --> 00:21:17,679 Speaker 1: you wish to convey. Also, the Tone Analyzer demo used 329 00:21:17,840 --> 00:21:22,560 Speaker 1: the business letter format as the means of comparison, so, 330 00:21:23,040 --> 00:21:27,920 Speaker 1: in other words, we compared Jane Austen to a business letter. Presumably, 331 00:21:27,960 --> 00:21:30,680 Speaker 1: if you were to use a full version of the analyzer, 332 00:21:30,760 --> 00:21:33,360 Speaker 1: not just the demo version, you would have other options 333 00:21:33,359 --> 00:21:37,520 Speaker 1: so you could compare it with other models, not just 334 00:21:37,560 --> 00:21:42,960 Speaker 1: a business letter. Joe McCormick. He included an excerpt from 335 00:21:43,560 --> 00:21:47,840 Speaker 1: Dostoyevsky's Notes from Underground. That excerpt was, I could not 336 00:21:47,960 --> 00:21:51,960 Speaker 1: become anything, neither good nor bad, neither a scoundrel nor 337 00:21:52,000 --> 00:21:55,520 Speaker 1: an honest man, neither a hero nor an insect. And 338 00:21:55,560 --> 00:21:58,400 Speaker 1: now I am eking out my days in my corner, 339 00:21:58,760 --> 00:22:03,480 Speaker 1: taunting myself with the bitter and entirely useless constellation that 340 00:22:03,600 --> 00:22:07,959 Speaker 1: an intelligent man cannot seriously become anything, that only a 341 00:22:08,000 --> 00:22:12,320 Speaker 1: fool can become something. The feedback was that the emotional 342 00:22:12,359 --> 00:22:19,120 Speaker 1: tone had anger at cheerfulness at so happy anger negative at. 343 00:22:20,119 --> 00:22:25,840 Speaker 1: The social tone was agreeable zero percent conscientious, zero percent open. 344 00:22:26,320 --> 00:22:31,320 Speaker 1: The writing tone was analytical, zero percent confident and tentative. 345 00:22:31,920 --> 00:22:34,479 Speaker 1: Joe would actually end up highlighting some of the words 346 00:22:34,960 --> 00:22:37,960 Speaker 1: to find out which words were the ones that ended 347 00:22:38,040 --> 00:22:44,080 Speaker 1: up giving that cheerfulness result. Those four words were good, honest, hero, 348 00:22:44,320 --> 00:22:51,200 Speaker 1: and intelligent and that kind of are that That's important 349 00:22:51,440 --> 00:22:55,600 Speaker 1: because those words, the way they are used, uh in 350 00:22:55,640 --> 00:22:59,800 Speaker 1: that passage are not used in a positive sense. They 351 00:23:00,000 --> 00:23:05,159 Speaker 1: are positive words, but they're meant to show kind of 352 00:23:05,200 --> 00:23:11,440 Speaker 1: a negation there not and not an assertion. So that 353 00:23:11,520 --> 00:23:14,880 Speaker 1: really highlights a big problem in this tone analyzer, which 354 00:23:14,920 --> 00:23:20,879 Speaker 1: is that it's tagging these words individually without context. So 355 00:23:20,920 --> 00:23:24,840 Speaker 1: if I wrote the phrase I am not glad, it 356 00:23:24,880 --> 00:23:27,680 Speaker 1: would tag the word glad and say that's a cheerful word. 357 00:23:28,359 --> 00:23:32,040 Speaker 1: But I said I am not glad. You if I 358 00:23:32,119 --> 00:23:35,120 Speaker 1: told you I am not glad, you would not think, oh, well, 359 00:23:35,119 --> 00:23:37,080 Speaker 1: that's a cheerful thing to say or a positive thing 360 00:23:37,119 --> 00:23:40,720 Speaker 1: to say. But according to the tone analyzer, it would 361 00:23:40,760 --> 00:23:44,040 Speaker 1: come across as a cheerful statement because it had tagged 362 00:23:44,080 --> 00:23:46,280 Speaker 1: that word as as being cheerful. In the other words 363 00:23:46,520 --> 00:23:50,040 Speaker 1: are not that strong, They don't They don't warrant being 364 00:23:50,040 --> 00:23:54,439 Speaker 1: tagged in a way like that. Now, over time, we 365 00:23:54,520 --> 00:23:57,520 Speaker 1: might have a tone analyzer that can actually take context 366 00:23:57,760 --> 00:24:02,040 Speaker 1: into account, and then you would learn a lot more 367 00:24:02,080 --> 00:24:05,840 Speaker 1: about the actual meaning behind a phrase. It would be 368 00:24:05,880 --> 00:24:08,679 Speaker 1: more than just tone. So if you were trying to 369 00:24:08,680 --> 00:24:13,960 Speaker 1: get across tone by using more complicated and subtle word 370 00:24:14,040 --> 00:24:19,160 Speaker 1: choice where you're sort of being kind of uh poetic 371 00:24:19,560 --> 00:24:23,680 Speaker 1: in your expression, you're trying to get across a feeling 372 00:24:24,160 --> 00:24:29,439 Speaker 1: by using irony or sarcasm, then a tone analyzer like 373 00:24:29,480 --> 00:24:31,879 Speaker 1: this would totally miss it because it would just be 374 00:24:31,920 --> 00:24:36,000 Speaker 1: counting the hits and not understanding the usage. There the 375 00:24:36,080 --> 00:24:40,480 Speaker 1: hidden meeting the word play, so that is going to 376 00:24:40,520 --> 00:24:45,760 Speaker 1: be a real challenge. So it's kind of another interesting 377 00:24:45,880 --> 00:24:48,120 Speaker 1: use of IBMS Watson. There are a lot of other 378 00:24:48,119 --> 00:24:50,600 Speaker 1: ones that we could talk about, like Chef Watson, which 379 00:24:50,680 --> 00:24:54,840 Speaker 1: was my favorite. Chef Watson would generate new recipes based 380 00:24:54,920 --> 00:24:57,280 Speaker 1: upon ingredients that you would tell it that you had 381 00:24:57,320 --> 00:25:01,600 Speaker 1: on hand, and it wouldn't it wouldn't go and reference 382 00:25:01,840 --> 00:25:05,520 Speaker 1: old recipes and pull one up for you. Instead, it 383 00:25:05,520 --> 00:25:09,200 Speaker 1: would make flavor profiles based upon all the different combinations 384 00:25:09,240 --> 00:25:11,800 Speaker 1: of food that were found in various recipe books and 385 00:25:11,840 --> 00:25:14,520 Speaker 1: generate a brand new recipe for you right there on 386 00:25:14,560 --> 00:25:18,280 Speaker 1: the spot. And sometimes they were whacka doodle crazy, y'all. 387 00:25:18,760 --> 00:25:21,959 Speaker 1: So in a way, you could say that Chef Watson 388 00:25:22,000 --> 00:25:25,320 Speaker 1: was another another way of seeing how I b M 389 00:25:25,359 --> 00:25:28,119 Speaker 1: S Watson has a lot of promise, but it requires 390 00:25:28,200 --> 00:25:32,240 Speaker 1: a ton of work on the app level in order 391 00:25:32,280 --> 00:25:35,760 Speaker 1: to leverage it and make actual practical use out of it. 392 00:25:36,160 --> 00:25:40,560 Speaker 1: I have more to say about computers detecting sarcasm, but 393 00:25:40,680 --> 00:25:52,960 Speaker 1: first let's take a quick word from our sponsor. So 394 00:25:53,400 --> 00:25:57,720 Speaker 1: back in two there were some researchers at the Hebrew 395 00:25:57,840 --> 00:26:02,760 Speaker 1: University in Israel who designed a system called the Semi 396 00:26:02,800 --> 00:26:10,080 Speaker 1: Supervised Algorithm for Sarcasm Identification or SAZI, and they used 397 00:26:10,160 --> 00:26:15,239 Speaker 1: SAZI to analyze collections of nearly six million tweets and 398 00:26:15,359 --> 00:26:20,320 Speaker 1: also around sixty six thousand product reviews from Amazon. They 399 00:26:20,400 --> 00:26:25,879 Speaker 1: wanted to find rich treasure troves of sarcasm that turns 400 00:26:25,880 --> 00:26:30,159 Speaker 1: out reviews and tweets. They fit the bill sarcasm is. 401 00:26:30,760 --> 00:26:35,600 Speaker 1: Really it's typically conveyed in some vocal tone, right and 402 00:26:35,680 --> 00:26:40,760 Speaker 1: nonverbal cues. So you have to first go someplace where 403 00:26:40,800 --> 00:26:44,520 Speaker 1: sarcasm is is rampant in text form to be able 404 00:26:44,560 --> 00:26:49,679 Speaker 1: to really fine tune how you can identify sarcasm versus 405 00:26:49,720 --> 00:26:52,600 Speaker 1: something that's meant exactly the way it's written on the 406 00:26:52,640 --> 00:26:57,920 Speaker 1: surface level. So they started to map out the various 407 00:26:58,040 --> 00:27:03,040 Speaker 1: features that were common in sarcastic comments online. So they 408 00:27:03,040 --> 00:27:06,080 Speaker 1: were looking for things like hyperbolic words and if you're 409 00:27:06,160 --> 00:27:09,640 Speaker 1: using a lot of exaggeration, that could be a key. 410 00:27:10,200 --> 00:27:14,600 Speaker 1: Excessive punctuation was another one, especially ellipses, which I tend 411 00:27:14,640 --> 00:27:17,080 Speaker 1: to use a lot, though I don't know if I 412 00:27:17,160 --> 00:27:19,159 Speaker 1: use it so much for sarcasm as I do for 413 00:27:19,359 --> 00:27:22,720 Speaker 1: just timing purposes. To indicate this is the beat I 414 00:27:22,760 --> 00:27:25,439 Speaker 1: would take if I were saying this out loud, I 415 00:27:25,480 --> 00:27:29,560 Speaker 1: guess that's just as irritating. Though, also how straightforward is 416 00:27:29,560 --> 00:27:32,840 Speaker 1: the Senate structure? And they gave it examples of sarcasm. 417 00:27:33,040 --> 00:27:37,719 Speaker 1: They fed it tweets that were tagged hashtag sarcasm, so 418 00:27:37,840 --> 00:27:42,400 Speaker 1: that the machine quote unquote knew that that was already 419 00:27:42,440 --> 00:27:45,399 Speaker 1: a sarcastic tweet and could start to analyze it and 420 00:27:45,480 --> 00:27:48,760 Speaker 1: build out a model for what sarcasm is. They also 421 00:27:48,800 --> 00:27:51,720 Speaker 1: fed at a bunch of one star Amazon reviews that 422 00:27:51,800 --> 00:27:55,880 Speaker 1: had been judged to be sarcastic by a panel consisting 423 00:27:55,920 --> 00:28:00,639 Speaker 1: of fifteen human beings, and the system was told it 424 00:28:00,720 --> 00:28:04,440 Speaker 1: had to rate sentences on a scale of one to five, 425 00:28:04,800 --> 00:28:09,120 Speaker 1: One being not sarcastic, they mean exactly what the Senate says, 426 00:28:09,560 --> 00:28:13,560 Speaker 1: five being holy cow, this person should write for the Onion. 427 00:28:13,920 --> 00:28:20,200 Speaker 1: This is incredibly sarcastic. SAZI could identify sarcastic Amazon reviews 428 00:28:20,720 --> 00:28:26,200 Speaker 1: with precision. Not bad, But when it came to Twitter 429 00:28:26,480 --> 00:28:31,040 Speaker 1: it did even better. I think, probably because there had 430 00:28:31,080 --> 00:28:33,560 Speaker 1: to be very short messages on Twitter. This was before 431 00:28:33,600 --> 00:28:36,960 Speaker 1: Twitter had even expanded to characters, so it was still 432 00:28:37,000 --> 00:28:40,240 Speaker 1: back in the one character days. The precision rate for 433 00:28:40,360 --> 00:28:45,200 Speaker 1: SAZI for Twitter was so it was really good at 434 00:28:45,280 --> 00:28:49,040 Speaker 1: detecting straightforward sarcasm, the kind that a lot of people 435 00:28:49,040 --> 00:28:52,160 Speaker 1: on Twitter use because you have limited space so you 436 00:28:52,200 --> 00:28:55,120 Speaker 1: can't really set it up in a more complex way. 437 00:28:55,160 --> 00:29:01,840 Speaker 1: But it was also uh more prone to judging things 438 00:29:01,960 --> 00:29:05,600 Speaker 1: as false negative evaluations rather than false positives. In other words, 439 00:29:05,960 --> 00:29:09,920 Speaker 1: it was more likely to look at a negative sarcastic 440 00:29:10,440 --> 00:29:13,200 Speaker 1: message and say that's not sarcastic than it was to 441 00:29:13,280 --> 00:29:16,640 Speaker 1: look at a straightforward message and say, no, that is sarcastic. 442 00:29:17,320 --> 00:29:20,920 Speaker 1: So that was kind of interesting back to Watson. Another 443 00:29:22,160 --> 00:29:25,480 Speaker 1: use of Watson came out of the Milk and Institute 444 00:29:25,480 --> 00:29:30,800 Speaker 1: Global Conference at IBM showed off some research that it 445 00:29:30,880 --> 00:29:34,280 Speaker 1: had been working on internally, and it was calling this 446 00:29:34,360 --> 00:29:40,000 Speaker 1: research debating Technologies. This was a project in which IBM 447 00:29:40,120 --> 00:29:42,480 Speaker 1: was trying to see if they could feed a computer 448 00:29:42,760 --> 00:29:48,400 Speaker 1: raw information, have the computer synthesize the information, understand that 449 00:29:48,520 --> 00:29:53,440 Speaker 1: information at least on a computational level, and then create 450 00:29:54,520 --> 00:29:59,880 Speaker 1: a a debating strategy for both pros and cons they 451 00:30:00,120 --> 00:30:02,960 Speaker 1: on that information. So it would take a huge amount 452 00:30:03,160 --> 00:30:08,360 Speaker 1: of content like all of Wikipedia, for example, and then 453 00:30:08,480 --> 00:30:11,440 Speaker 1: on any given subject that would be covered in Wikipedia. 454 00:30:11,480 --> 00:30:14,600 Speaker 1: It would be asked form an argument that is in 455 00:30:14,680 --> 00:30:19,520 Speaker 1: favor of or is against a concept, whatever that concept 456 00:30:19,560 --> 00:30:22,560 Speaker 1: might be. John Kelly of IBM showed off in a 457 00:30:22,640 --> 00:30:25,600 Speaker 1: demo how the tool could be used to predict pro 458 00:30:25,840 --> 00:30:29,760 Speaker 1: or con arguments about a subject based on a body 459 00:30:29,960 --> 00:30:34,680 Speaker 1: of information. So you might be able to use this 460 00:30:34,920 --> 00:30:41,360 Speaker 1: technology in order to anticipate what an opposing person might 461 00:30:41,480 --> 00:30:44,440 Speaker 1: say on any given subject. Let's say that you are 462 00:30:44,560 --> 00:30:49,400 Speaker 1: getting ready to debate a topic. You might feed that 463 00:30:49,720 --> 00:30:53,600 Speaker 1: information to a computer system using this Watson platform. You 464 00:30:53,680 --> 00:30:56,480 Speaker 1: might feed in a ton of information, and then you 465 00:30:56,560 --> 00:31:00,760 Speaker 1: might say, who is a man and someone who is 466 00:31:00,800 --> 00:31:05,840 Speaker 1: against this particular topic, whatever it might be. Uh. Let's 467 00:31:05,840 --> 00:31:10,880 Speaker 1: say it's it's it's renewable energy, and the uh the 468 00:31:11,120 --> 00:31:13,840 Speaker 1: efficiency of solar panels, whether or not it makes sense 469 00:31:13,880 --> 00:31:17,040 Speaker 1: to invest in solar panels. Let's say that your stance 470 00:31:17,320 --> 00:31:20,400 Speaker 1: is that you have to argue for solar panels. You 471 00:31:20,480 --> 00:31:23,640 Speaker 1: might say, what would someone who wants to argue against 472 00:31:23,920 --> 00:31:29,000 Speaker 1: solar panels say? And then Watson would analyze this information 473 00:31:29,560 --> 00:31:33,880 Speaker 1: and return to you what it thinks would be an 474 00:31:33,960 --> 00:31:39,840 Speaker 1: argument someone would use to support that that stance, and 475 00:31:39,880 --> 00:31:42,120 Speaker 1: then you could prepare for that, which would be an 476 00:31:42,160 --> 00:31:44,120 Speaker 1: incredible tool. I mean, you could think of this as 477 00:31:44,160 --> 00:31:46,760 Speaker 1: for political debates. It would be amazing. You could think 478 00:31:46,760 --> 00:31:49,760 Speaker 1: of how you might want to prepare so that you 479 00:31:49,840 --> 00:31:52,960 Speaker 1: can argue intelligently against an opponent, and you can already 480 00:31:53,000 --> 00:31:55,880 Speaker 1: anticipate what that opponent is going to say, because you 481 00:31:55,880 --> 00:31:58,680 Speaker 1: know their general stance on a topic, but you might 482 00:31:58,720 --> 00:32:02,080 Speaker 1: not know what tactic they might use to support that stance. 483 00:32:03,000 --> 00:32:05,920 Speaker 1: Maybe politics isn't a great choice because that's not always 484 00:32:05,960 --> 00:32:09,680 Speaker 1: in the realm of rationality. That often falls into a 485 00:32:11,040 --> 00:32:15,440 Speaker 1: call toward emotional response rather than rational response. That's more 486 00:32:15,480 --> 00:32:20,120 Speaker 1: of a a commentary on politics in general, regardless of 487 00:32:20,160 --> 00:32:23,240 Speaker 1: what side you might be on, all sides do this anyway. 488 00:32:23,600 --> 00:32:27,680 Speaker 1: He actually showed at this demo a different example. He said, 489 00:32:28,040 --> 00:32:30,400 Speaker 1: what if you were to take the sale of violent 490 00:32:30,520 --> 00:32:34,240 Speaker 1: video games to minors should be banned. That's the topic, 491 00:32:35,160 --> 00:32:37,840 Speaker 1: and that the computer would then go through all the 492 00:32:37,840 --> 00:32:41,480 Speaker 1: information and had access to. It would end up sorting 493 00:32:41,480 --> 00:32:44,520 Speaker 1: out all the parts that were relevant to the discussion, 494 00:32:45,120 --> 00:32:47,120 Speaker 1: so it just put those aside and that would become 495 00:32:47,200 --> 00:32:51,360 Speaker 1: the core of the data it would reference. I would 496 00:32:51,360 --> 00:32:54,080 Speaker 1: then go through and identify basic statements as either being 497 00:32:54,480 --> 00:32:59,880 Speaker 1: a pro stance of banning violent video games to my 498 00:33:00,000 --> 00:33:04,120 Speaker 1: nes or a constance for that saying no, we should 499 00:33:04,160 --> 00:33:08,160 Speaker 1: be able to sell violent video games to minors. The 500 00:33:08,160 --> 00:33:11,720 Speaker 1: tools scanned four million articles, it returned the top ten 501 00:33:11,920 --> 00:33:14,800 Speaker 1: articles that were determined to be the most relevant to 502 00:33:15,040 --> 00:33:19,760 Speaker 1: that particular debate, and it scanned approximately three thousand sentences 503 00:33:20,720 --> 00:33:24,120 Speaker 1: from from top to bottom, and it then identified sentences 504 00:33:24,120 --> 00:33:28,640 Speaker 1: that contained candidate claims that would be statements that would 505 00:33:28,680 --> 00:33:32,280 Speaker 1: either be interpreted as being pro or con for the stance. 506 00:33:32,920 --> 00:33:35,720 Speaker 1: Then it identified the parameters of those claims. Then it 507 00:33:35,760 --> 00:33:38,960 Speaker 1: assessed the claims for the pro and con polarity, then 508 00:33:39,000 --> 00:33:42,760 Speaker 1: constructed a sample pro or con statement. And the statements 509 00:33:42,800 --> 00:33:45,760 Speaker 1: in the demo were kind of interesting. And since the 510 00:33:45,760 --> 00:33:50,440 Speaker 1: computer is constructing arguments based upon what people have already written, 511 00:33:51,080 --> 00:33:53,640 Speaker 1: it would reflect a lot of vague statements that aren't 512 00:33:53,640 --> 00:33:56,200 Speaker 1: a firm stance. So, in other words, like it couldn't 513 00:33:56,200 --> 00:33:59,560 Speaker 1: take a bunch of stuff that was written that it 514 00:33:59,680 --> 00:34:03,640 Speaker 1: sell did not take either a pro or constance, and 515 00:34:03,680 --> 00:34:07,080 Speaker 1: then transformed that magically into the perfect pro stance or 516 00:34:07,120 --> 00:34:10,640 Speaker 1: the perfect constance. Uh, it's dependent upon the words that 517 00:34:10,760 --> 00:34:14,440 Speaker 1: human beings have already written, so it could not magically 518 00:34:14,440 --> 00:34:17,440 Speaker 1: come up with a killer argument if the data that 519 00:34:17,480 --> 00:34:21,279 Speaker 1: had been written about this subject didn't come down on 520 00:34:21,480 --> 00:34:26,640 Speaker 1: a firm stance one way or the other. Um, the 521 00:34:26,680 --> 00:34:29,319 Speaker 1: point of the demonstration wasn't to create a tool that 522 00:34:29,360 --> 00:34:34,680 Speaker 1: could either troll people or counter trolls. It was to 523 00:34:34,719 --> 00:34:37,279 Speaker 1: show that a computer could be useful to aid in 524 00:34:37,320 --> 00:34:41,760 Speaker 1: the reasoning process when you're making a critical decision. Again, 525 00:34:41,800 --> 00:34:44,360 Speaker 1: to go back to that medical example, it could be 526 00:34:44,440 --> 00:34:48,600 Speaker 1: used to help a doctor determine which diagnosis is the 527 00:34:48,640 --> 00:34:51,800 Speaker 1: most likely to be accurate for a patient, what what 528 00:34:52,560 --> 00:34:55,720 Speaker 1: course of treatment might be the most helpful for that patient, 529 00:34:56,520 --> 00:35:01,080 Speaker 1: and thus it could have real practic coal use outside 530 00:35:01,120 --> 00:35:07,799 Speaker 1: of this more esoteric, interesting uh debate news. Now, will 531 00:35:07,840 --> 00:35:12,520 Speaker 1: we see computers in the future able to detect sarcasm 532 00:35:12,560 --> 00:35:16,279 Speaker 1: just as easily as your typical human being can when 533 00:35:16,320 --> 00:35:21,960 Speaker 1: given the right circumstances. And I use the word typical reluctantly, 534 00:35:22,320 --> 00:35:25,319 Speaker 1: but you get what I mean. I don't know. It's 535 00:35:25,360 --> 00:35:27,680 Speaker 1: gonna take some time. It takes an awful lot of 536 00:35:27,680 --> 00:35:30,800 Speaker 1: processing power too. You have to remember that for these 537 00:35:30,920 --> 00:35:34,719 Speaker 1: neural networks systems, the ones that are running these these 538 00:35:34,800 --> 00:35:39,360 Speaker 1: various platforms and programs and strategies. They take up a 539 00:35:39,400 --> 00:35:46,640 Speaker 1: lot of processing power because our brains have billion neurons 540 00:35:46,640 --> 00:35:50,680 Speaker 1: in them, so we have a very sophisticated supercomputer sitting 541 00:35:50,719 --> 00:35:55,920 Speaker 1: in our heads. Moreover, our brains are insanely energy efficient. 542 00:35:56,040 --> 00:35:58,960 Speaker 1: They require about the equivalent of twenty watts of power. 543 00:36:00,000 --> 00:36:03,600 Speaker 1: A supercomputer needs a lot more power than that. So 544 00:36:04,400 --> 00:36:07,919 Speaker 1: while we're seeing advances in this, it requires so much 545 00:36:07,960 --> 00:36:11,239 Speaker 1: processing power, so much energy, it is not a practical 546 00:36:12,120 --> 00:36:16,440 Speaker 1: approach to most forms of computing, at least from a 547 00:36:16,480 --> 00:36:20,640 Speaker 1: consumer standpoint. You might see a future where the sort 548 00:36:20,680 --> 00:36:23,239 Speaker 1: of stuff is all in the cloud and then we 549 00:36:23,280 --> 00:36:26,920 Speaker 1: can access it through an app or a program or whatever. 550 00:36:27,400 --> 00:36:29,959 Speaker 1: That way, you don't have to have a supercomputer sitting 551 00:36:30,000 --> 00:36:32,920 Speaker 1: on your desk in order to tap into those, uh, 552 00:36:33,040 --> 00:36:35,360 Speaker 1: those capabilities, but you have to have an Internet connection, 553 00:36:35,400 --> 00:36:39,680 Speaker 1: which most of us these days tend to have fairly frequently. 554 00:36:39,760 --> 00:36:41,040 Speaker 1: I mean, there are a lot of people out there 555 00:36:41,040 --> 00:36:44,680 Speaker 1: who at this point have had a persistent Internet connection 556 00:36:44,800 --> 00:36:48,280 Speaker 1: for pretty much their whole lives, which blows my mind. 557 00:36:48,840 --> 00:36:50,319 Speaker 1: But that's the kind of world we'd have to live 558 00:36:50,360 --> 00:36:52,759 Speaker 1: in in order to really take advantage of this at 559 00:36:52,800 --> 00:36:55,839 Speaker 1: least in the near term. I don't know if we're 560 00:36:55,840 --> 00:36:59,040 Speaker 1: ever going to see a computer that can analyze, say, 561 00:36:59,040 --> 00:37:02,799 Speaker 1: an article from the Onion and not only point out 562 00:37:02,800 --> 00:37:06,279 Speaker 1: that it's being sarcastic or ironic, but also point out 563 00:37:06,280 --> 00:37:08,759 Speaker 1: why it's funny. I think at one point, when you 564 00:37:08,800 --> 00:37:12,440 Speaker 1: start analyzing comedy, there gets to be a level where 565 00:37:12,480 --> 00:37:15,360 Speaker 1: nothing is ever funny ever again, but it is a 566 00:37:15,400 --> 00:37:20,520 Speaker 1: really interesting problem. So that's whether that's that's this look 567 00:37:20,560 --> 00:37:25,840 Speaker 1: back on if AI is ever going to understand sarcasm. Well, guys, 568 00:37:25,840 --> 00:37:29,200 Speaker 1: I hope you enjoyed that classic episode of tech stuff. 569 00:37:29,239 --> 00:37:31,839 Speaker 1: I guess I guess two years old isn't old enough 570 00:37:31,880 --> 00:37:36,280 Speaker 1: to be classic. That uh that that only somewhat less 571 00:37:36,280 --> 00:37:40,400 Speaker 1: than fresh episode of text stuff about artificial intelligence and 572 00:37:40,480 --> 00:37:44,920 Speaker 1: sarcasm and things of that nature. I am constantly impressed 573 00:37:45,200 --> 00:37:49,560 Speaker 1: with how artificial intelligence is advancing year over year. But 574 00:37:50,160 --> 00:37:52,600 Speaker 1: when you look at what it means to be human 575 00:37:53,000 --> 00:37:56,280 Speaker 1: and the ways that we humans interact with one another, 576 00:37:56,480 --> 00:38:00,480 Speaker 1: and the ways that we can communicate complicated three things, 577 00:38:00,560 --> 00:38:03,960 Speaker 1: sometimes just through you know, subtle methods that are not 578 00:38:04,520 --> 00:38:09,520 Speaker 1: overt or or you know, directly spoken, it reminds us 579 00:38:09,560 --> 00:38:11,880 Speaker 1: that machines have got a long way to go in 580 00:38:12,000 --> 00:38:14,640 Speaker 1: order to really grasp what it is to be human, 581 00:38:14,920 --> 00:38:19,279 Speaker 1: So unless you're Commander Data, you're probably struggling a bit. 582 00:38:20,000 --> 00:38:22,040 Speaker 1: So I hope you guys enjoyed this. If you have 583 00:38:22,080 --> 00:38:24,799 Speaker 1: suggestions for future episodes of tech Stuff, I've got a 584 00:38:24,800 --> 00:38:28,560 Speaker 1: few episodes based on listener suggestions coming up soon. But 585 00:38:28,680 --> 00:38:31,440 Speaker 1: if you want to get your suggestions in tweet me. 586 00:38:31,920 --> 00:38:35,880 Speaker 1: The Twitter handle is text stuff H s W and 587 00:38:35,920 --> 00:38:44,040 Speaker 1: I'll talk to you again really soon. Text Stuff is 588 00:38:44,080 --> 00:38:47,200 Speaker 1: an I Heart Radio production. For more podcasts from my 589 00:38:47,360 --> 00:38:50,960 Speaker 1: Heart Radio, visit the i Heart Radio app, Apple Podcasts, 590 00:38:51,080 --> 00:38:53,080 Speaker 1: or wherever you listen to your favorite shows.