1 00:00:04,120 --> 00:00:07,160 Speaker 1: Get in touch with technology with tech Stuff from how 2 00:00:07,200 --> 00:00:14,080 Speaker 1: stuff Works dot com. Hey there, and welcome to tech Stuff. 3 00:00:14,120 --> 00:00:16,960 Speaker 1: I'm your host, Jonathan Strickland. I'm an executive producer with 4 00:00:17,000 --> 00:00:20,560 Speaker 1: How Stuff Works and I love all things tech. Today, 5 00:00:21,079 --> 00:00:24,000 Speaker 1: I want to talk to you about an interesting topic 6 00:00:24,160 --> 00:00:26,599 Speaker 1: that I got to explore a couple of years ago 7 00:00:27,120 --> 00:00:31,760 Speaker 1: with Joe McCormick and Lauren fogobaum As we debated the 8 00:00:31,840 --> 00:00:37,800 Speaker 1: possibilities of computers learning how to understand sarcasm. We did 9 00:00:37,840 --> 00:00:41,160 Speaker 1: it for a podcast called Forward Thinking, which was around 10 00:00:41,200 --> 00:00:42,920 Speaker 1: for a couple of years. It was a lot of 11 00:00:42,920 --> 00:00:46,040 Speaker 1: fun to work on that that show is over, but 12 00:00:46,080 --> 00:00:48,640 Speaker 1: I thought I would revisit the topic and talk about 13 00:00:48,680 --> 00:00:52,120 Speaker 1: it for you guys and kind of go over what 14 00:00:52,280 --> 00:00:54,800 Speaker 1: would it take to have a computer that could actually 15 00:00:54,880 --> 00:00:59,400 Speaker 1: understand when someone's being sarcastic. Now to understand why this 16 00:00:59,440 --> 00:01:02,360 Speaker 1: is a big d it helps to have a refresher 17 00:01:02,400 --> 00:01:05,679 Speaker 1: course on how computers process information. And I know I 18 00:01:05,720 --> 00:01:08,560 Speaker 1: talked about this a lot, but I still think it's 19 00:01:08,560 --> 00:01:11,200 Speaker 1: important to cover the basics when you want to talk 20 00:01:11,240 --> 00:01:14,840 Speaker 1: about something as advanced as being able to detect and 21 00:01:15,040 --> 00:01:21,200 Speaker 1: understand sarcasm. So computers understand machine code or assembly language. 22 00:01:21,480 --> 00:01:25,320 Speaker 1: This is a language that corresponds with the actual physical 23 00:01:25,600 --> 00:01:30,319 Speaker 1: architecture of the computers, so the way the computer is built, 24 00:01:30,680 --> 00:01:33,880 Speaker 1: that's how this language interacts. It's it's essentially how the 25 00:01:33,959 --> 00:01:39,119 Speaker 1: physical components of the computer are able to handle electric 26 00:01:39,200 --> 00:01:45,560 Speaker 1: current or voltage differences in order to process information, and 27 00:01:45,880 --> 00:01:51,680 Speaker 1: computers can interpret this and execute upon this language very quickly. 28 00:01:52,240 --> 00:01:56,160 Speaker 1: It is the basic language of those physical components. However, 29 00:01:57,000 --> 00:02:00,880 Speaker 1: it is almost impossible for human to work with this, 30 00:02:01,200 --> 00:02:04,040 Speaker 1: at least on a way that is at all efficient, 31 00:02:04,480 --> 00:02:10,800 Speaker 1: because it ultimately for most computers boils down to binary language, right, 32 00:02:11,000 --> 00:02:16,120 Speaker 1: zeros and ones, So you see a huge block of 33 00:02:16,200 --> 00:02:18,799 Speaker 1: zeros and ones, and unless you are neo from the matrix, 34 00:02:18,840 --> 00:02:22,920 Speaker 1: it means nothing to you. So we speak in natural 35 00:02:23,240 --> 00:02:27,480 Speaker 1: language to one another. Natural language, however, is filled with 36 00:02:27,520 --> 00:02:31,640 Speaker 1: a lot of components that make it very very challenging 37 00:02:31,680 --> 00:02:36,160 Speaker 1: for machines to interpret, like ambiguity, or there might be 38 00:02:36,200 --> 00:02:39,200 Speaker 1: double meanings in a phrase and you may mean both 39 00:02:39,280 --> 00:02:43,960 Speaker 1: meanings at the same time, and that is too complicated 40 00:02:44,000 --> 00:02:46,200 Speaker 1: for most machines to be able to process. They just 41 00:02:46,240 --> 00:02:50,560 Speaker 1: can't deal with that. So to bridge the gap between 42 00:02:50,760 --> 00:02:54,480 Speaker 1: the way we humans communicate and the way that computers 43 00:02:54,600 --> 00:03:00,440 Speaker 1: process language. We have created programming languages and compilers. Now, 44 00:03:00,800 --> 00:03:04,760 Speaker 1: programming languages fall into two broad categories. It's more like 45 00:03:05,080 --> 00:03:07,920 Speaker 1: a spectrum, and you could be further on one end 46 00:03:08,000 --> 00:03:11,320 Speaker 1: than the other, and we typically call them high level 47 00:03:11,560 --> 00:03:15,960 Speaker 1: programming languages and low level programming languages. The lower the 48 00:03:16,120 --> 00:03:19,920 Speaker 1: level of programming language, the closer it is to machine code, 49 00:03:20,560 --> 00:03:23,399 Speaker 1: and the easier it is for a computer to understand, 50 00:03:23,800 --> 00:03:26,040 Speaker 1: but the harder it is to work with if you 51 00:03:26,080 --> 00:03:29,200 Speaker 1: happen to be, you know, a human being. High level 52 00:03:29,240 --> 00:03:33,480 Speaker 1: programming languages are easier for humans to understand. Now, if 53 00:03:33,520 --> 00:03:36,960 Speaker 1: you have never taken any courses in programming and you're 54 00:03:37,000 --> 00:03:41,040 Speaker 1: looking at a page of code, it can seem indecipherable 55 00:03:41,040 --> 00:03:46,360 Speaker 1: to you. It is just meaningless strings of characters. But 56 00:03:47,240 --> 00:03:50,720 Speaker 1: once you learn the rules of that programming language, how 57 00:03:50,800 --> 00:03:54,800 Speaker 1: you construct an instruction and a series of instructions, how 58 00:03:54,840 --> 00:03:57,640 Speaker 1: you go from one instruction to the next. Once you 59 00:03:57,720 --> 00:04:00,920 Speaker 1: understand the rules, it actually becomes quite easy to use 60 00:04:01,160 --> 00:04:03,200 Speaker 1: in the grand scheme of things, much more easy than 61 00:04:03,280 --> 00:04:06,880 Speaker 1: machine language would be. But again, the problem here is 62 00:04:06,920 --> 00:04:11,960 Speaker 1: that computers don't understand programming languages, not natively. Even though 63 00:04:12,480 --> 00:04:15,280 Speaker 1: this is not exactly the same as human natural language, 64 00:04:15,280 --> 00:04:18,039 Speaker 1: it's also not the same as machine language. That's why 65 00:04:18,040 --> 00:04:23,719 Speaker 1: you need compilers. A compiler is essentially a translator. It 66 00:04:23,880 --> 00:04:28,479 Speaker 1: takes this high level programming language or higher level anyway, 67 00:04:28,560 --> 00:04:32,080 Speaker 1: and then converts it into a machine readable language for 68 00:04:32,120 --> 00:04:35,400 Speaker 1: the computer to actually execute upon. And this is all 69 00:04:35,440 --> 00:04:39,080 Speaker 1: in the design of the programming languages and the compilers. 70 00:04:40,040 --> 00:04:44,039 Speaker 1: So this is the way that for decades we have 71 00:04:44,120 --> 00:04:46,760 Speaker 1: interacted with computers, when you're talking about it on a 72 00:04:46,839 --> 00:04:49,680 Speaker 1: on a direct level, not just executing a program, but 73 00:04:49,839 --> 00:04:54,599 Speaker 1: creating code, creating programs for computers to run. Over the 74 00:04:54,720 --> 00:04:58,960 Speaker 1: last few decades, we've had some very very smart people 75 00:04:59,520 --> 00:05:05,599 Speaker 1: working on natural language systems for machines which would allow 76 00:05:05,839 --> 00:05:12,560 Speaker 1: a computer to interpret natural language in a way that 77 00:05:12,560 --> 00:05:14,920 Speaker 1: would make some sort of sense and for the computer 78 00:05:14,960 --> 00:05:17,320 Speaker 1: to be able to act upon that language. And we've 79 00:05:17,360 --> 00:05:22,479 Speaker 1: seen this in plenty of examples recently. Most smartphones have 80 00:05:22,680 --> 00:05:26,560 Speaker 1: some sort of smart assistant. You have standalone products like 81 00:05:26,720 --> 00:05:31,000 Speaker 1: Amazon's Echo, you have Google Home, You've got tons of 82 00:05:31,080 --> 00:05:37,080 Speaker 1: devices that can interact with people. It can be activated 83 00:05:37,120 --> 00:05:39,800 Speaker 1: by typically an alert phrase, which I'm not going to 84 00:05:39,880 --> 00:05:41,680 Speaker 1: say because I don't want any of you guys to 85 00:05:41,720 --> 00:05:43,880 Speaker 1: have to deal with that. I know how irritating it 86 00:05:43,960 --> 00:05:47,400 Speaker 1: is when I'm watching a video and someone activates their 87 00:05:48,760 --> 00:05:52,920 Speaker 1: specific system and then mine begins to respond and all 88 00:05:52,920 --> 00:05:55,640 Speaker 1: my lights started going on and off because the people 89 00:05:55,640 --> 00:05:58,560 Speaker 1: on YouTube we're talking funny. I know how irritating that is. 90 00:05:58,600 --> 00:06:01,680 Speaker 1: But use that at debates and then you can speak 91 00:06:02,080 --> 00:06:06,400 Speaker 1: and typically you can say the same thing several different 92 00:06:06,400 --> 00:06:11,520 Speaker 1: ways and the device appears to understand you no matter 93 00:06:11,560 --> 00:06:14,279 Speaker 1: how you word it. And this is a real challenge 94 00:06:14,279 --> 00:06:17,120 Speaker 1: because we human beings can find lots of different ways 95 00:06:17,560 --> 00:06:20,360 Speaker 1: to say the same thing. For example, if I say 96 00:06:20,400 --> 00:06:23,560 Speaker 1: what is the weather today, it could be very similar 97 00:06:23,600 --> 00:06:25,640 Speaker 1: to if I if I ask a question, is it 98 00:06:25,720 --> 00:06:29,120 Speaker 1: going to rain today? Both of those are asking for 99 00:06:29,160 --> 00:06:32,560 Speaker 1: information about the weather, but are very different ways of 100 00:06:32,600 --> 00:06:36,760 Speaker 1: saying that. A good natural language recognition program will be 101 00:06:36,800 --> 00:06:42,360 Speaker 1: able to parse that information and then return the appropriate response. 102 00:06:43,600 --> 00:06:46,760 Speaker 1: This is not an easy thing to do. Typically it 103 00:06:46,800 --> 00:06:50,880 Speaker 1: involves creating a neural network structure, and I've talked about 104 00:06:50,960 --> 00:06:55,640 Speaker 1: artificial neural networks recently. That's a typically a network that 105 00:06:55,720 --> 00:07:01,440 Speaker 1: can accept multiple binary inputs, so either a zero or 106 00:07:01,520 --> 00:07:06,640 Speaker 1: a one input that represents something uh, some sort of yes, 107 00:07:06,720 --> 00:07:10,440 Speaker 1: no or on off kind of feature. It can accept 108 00:07:10,560 --> 00:07:14,760 Speaker 1: multiple multiple inputs of that nature, so multiple zeros or 109 00:07:14,840 --> 00:07:18,920 Speaker 1: ones that all factor into making a decision, and then 110 00:07:18,920 --> 00:07:22,720 Speaker 1: it has a waiting for each of those components, and 111 00:07:22,760 --> 00:07:26,400 Speaker 1: then it produces a single output that's also binary in nature, 112 00:07:26,440 --> 00:07:28,920 Speaker 1: either a zero one, and it passes that on to 113 00:07:29,240 --> 00:07:33,440 Speaker 1: other artificial neurons further down the chain. Sometimes that will 114 00:07:33,480 --> 00:07:37,080 Speaker 1: come back around and you have a recursive artificial neural network. 115 00:07:37,440 --> 00:07:42,920 Speaker 1: The goal here is for this process two ultimately result 116 00:07:43,760 --> 00:07:49,080 Speaker 1: in a response that is reasonably certain to meet the 117 00:07:49,120 --> 00:07:52,800 Speaker 1: requirements of the person asking the question. This tends to 118 00:07:52,800 --> 00:07:56,720 Speaker 1: be talked about in the realm of probabilities. We we 119 00:07:56,760 --> 00:08:00,280 Speaker 1: talked about how certain the machine is that the respons 120 00:08:00,400 --> 00:08:03,240 Speaker 1: is the appropriate one, and if it falls below a 121 00:08:03,280 --> 00:08:07,800 Speaker 1: certain threshold, then the machine would typically respond with I'm sorry, 122 00:08:07,840 --> 00:08:10,040 Speaker 1: I don't know what you're asking for, or something similar 123 00:08:10,080 --> 00:08:13,840 Speaker 1: to that. There are cases where you just get misinterpreted 124 00:08:13,960 --> 00:08:16,559 Speaker 1: and you'll get a response that does not reflect whatever 125 00:08:16,600 --> 00:08:18,760 Speaker 1: you ask. That's a little different. That's where the machine 126 00:08:18,760 --> 00:08:22,760 Speaker 1: has drawn a conclusion, has been reasonably certain that it 127 00:08:22,800 --> 00:08:24,680 Speaker 1: came to the right conclusion, it turns out it was 128 00:08:24,720 --> 00:08:29,240 Speaker 1: wrong the whole way. But that's the process. Now, when 129 00:08:29,280 --> 00:08:36,559 Speaker 1: it comes to sarcasm, that adds yet another layer of difficulty, 130 00:08:37,320 --> 00:08:42,120 Speaker 1: because now a machine isn't just parsing what you are saying. 131 00:08:42,520 --> 00:08:46,520 Speaker 1: It has to understand what you mean, the meaning of 132 00:08:46,559 --> 00:08:51,480 Speaker 1: your words and the meaning of the way you deliver them. 133 00:08:51,480 --> 00:08:54,120 Speaker 1: It could be different. So if I were to just 134 00:08:54,240 --> 00:08:59,360 Speaker 1: write out a phrase with no tone, no body language, uh, 135 00:08:59,600 --> 00:09:03,920 Speaker 1: not emphasizing any one word over another, it might be 136 00:09:04,040 --> 00:09:08,319 Speaker 1: very difficult to detect what my intent was. It may 137 00:09:08,360 --> 00:09:11,559 Speaker 1: seem like I'm being sincere, when in fact I'm being insincere. 138 00:09:11,840 --> 00:09:16,280 Speaker 1: For example, Uh, if I were to say that guy 139 00:09:16,400 --> 00:09:22,040 Speaker 1: is super tall, but I'm being sarcastic, then just in 140 00:09:22,080 --> 00:09:25,440 Speaker 1: that phrase the way I write it out, you would think, oh, well, 141 00:09:25,480 --> 00:09:29,959 Speaker 1: that person he's looking at must be super tall. How 142 00:09:30,000 --> 00:09:34,120 Speaker 1: do you recognize sarcasm? How can you detect that this 143 00:09:34,200 --> 00:09:37,280 Speaker 1: is in place and then understand what the meaning underneath 144 00:09:37,320 --> 00:09:41,760 Speaker 1: it is. One of the approaches that has been put 145 00:09:41,800 --> 00:09:48,480 Speaker 1: forward relates to IBM's Watson platform. Now. Watson first made 146 00:09:48,480 --> 00:09:52,440 Speaker 1: headlines back when it was a contestant on Jeopardy. It 147 00:09:52,720 --> 00:09:56,880 Speaker 1: went up against two former champions, including Ken Jennings, who 148 00:09:57,000 --> 00:10:00,240 Speaker 1: shows up on a house Stuff Works podcast. Anyway, Utson 149 00:10:00,280 --> 00:10:03,840 Speaker 1: went up against these two former champions and it was 150 00:10:03,920 --> 00:10:07,160 Speaker 1: able to interpret natural language. It had to in order 151 00:10:07,200 --> 00:10:09,120 Speaker 1: to play the game of Jeopardy. And for those who 152 00:10:09,200 --> 00:10:11,920 Speaker 1: do not know what Jeopardy is or they're not familiar 153 00:10:11,920 --> 00:10:15,120 Speaker 1: with the game show, Jeopardy is a game where you 154 00:10:15,160 --> 00:10:21,079 Speaker 1: are presented with categories of trivia and each category has 155 00:10:21,200 --> 00:10:27,679 Speaker 1: multiple uh questions or multiple entries in it, and they 156 00:10:27,800 --> 00:10:33,360 Speaker 1: range in dollar value, and the lower dollar value ones 157 00:10:33,400 --> 00:10:37,000 Speaker 1: are easier to answer than the higher dollar value ones, 158 00:10:38,120 --> 00:10:41,680 Speaker 1: and UH, you're Typically the way Jeopardy works is that 159 00:10:41,720 --> 00:10:44,600 Speaker 1: you're you're given quote unquote the answer and you have 160 00:10:44,679 --> 00:10:49,840 Speaker 1: to provide the question. So uh, if the answer were 161 00:10:51,360 --> 00:10:57,440 Speaker 1: this film that detailed the adventures of a young playwright 162 00:10:57,640 --> 00:11:01,920 Speaker 1: in sixteenth century England one picture, you would say, what 163 00:11:02,080 --> 00:11:06,240 Speaker 1: was Shakespeare in Love? So this computer is playing against 164 00:11:06,240 --> 00:11:08,920 Speaker 1: these two former champions. This was sort of an exhibition 165 00:11:09,480 --> 00:11:14,160 Speaker 1: series of games. It wasn't meant for uh, a competition 166 00:11:14,200 --> 00:11:16,480 Speaker 1: in the way the typical Jeopardy games were. There was 167 00:11:16,559 --> 00:11:19,960 Speaker 1: money on the line. It was an exhibition and Watson 168 00:11:20,000 --> 00:11:23,160 Speaker 1: won it beat both of the champions, and it did 169 00:11:23,160 --> 00:11:26,440 Speaker 1: what I was telling you. It it would analyze the 170 00:11:26,600 --> 00:11:30,719 Speaker 1: clue that was given, the answer that was given, it 171 00:11:30,760 --> 00:11:33,959 Speaker 1: would try and generate a question to correspond with that answer, 172 00:11:34,360 --> 00:11:37,480 Speaker 1: and only if the question met a certain threshold of 173 00:11:37,520 --> 00:11:40,600 Speaker 1: confidence with Watson buzz in. If it did not meet 174 00:11:40,960 --> 00:11:45,040 Speaker 1: that level of confidence, Watson would remain quiet. And most importantly, 175 00:11:45,320 --> 00:11:47,920 Speaker 1: Watson was not at all connected to the Internet. All 176 00:11:48,000 --> 00:11:53,640 Speaker 1: the information was contained within a massive series of servers 177 00:11:54,559 --> 00:11:57,080 Speaker 1: more than gosh, I can't even remember. There's a ton 178 00:11:57,160 --> 00:12:02,440 Speaker 1: of processors attached to it. Um so a very powerful machine, 179 00:12:03,520 --> 00:12:09,640 Speaker 1: but it still wasn't exactly able to detect sarcasm. It 180 00:12:09,720 --> 00:12:14,040 Speaker 1: could work with wordplay, and it could work with riddles, 181 00:12:14,040 --> 00:12:16,960 Speaker 1: so that was really impressive. But what it really did 182 00:12:17,000 --> 00:12:19,560 Speaker 1: was it gave IBM the opportunity to say, we have 183 00:12:19,720 --> 00:12:24,360 Speaker 1: this platform here, and we're welcoming developers to create applications 184 00:12:24,400 --> 00:12:28,160 Speaker 1: that tap into this platform and make use of this 185 00:12:28,880 --> 00:12:32,640 Speaker 1: in order to do interesting stuff with it. And IBM 186 00:12:32,720 --> 00:12:35,319 Speaker 1: was largely working with the medical industry at that point 187 00:12:35,360 --> 00:12:41,600 Speaker 1: to try and help doctors treat and diagnose patients, and 188 00:12:41,679 --> 00:12:43,760 Speaker 1: it was sort of computer guidance. It wasn't that you 189 00:12:43,840 --> 00:12:47,960 Speaker 1: had an automatic doctor, but rather the doctor had what 190 00:12:48,320 --> 00:12:53,480 Speaker 1: equates to a medical expert to confer with when trying 191 00:12:53,520 --> 00:12:56,760 Speaker 1: to determine why's the best course of action for a patient. 192 00:12:57,800 --> 00:13:01,120 Speaker 1: IBM put up an Application program m interface or API 193 00:13:01,640 --> 00:13:06,320 Speaker 1: and let developers create their own cognitive computing applications built 194 00:13:06,400 --> 00:13:10,600 Speaker 1: on top of Watson. One of those was called the 195 00:13:10,640 --> 00:13:14,680 Speaker 1: tone analyzer. It still exists back when we were doing 196 00:13:14,679 --> 00:13:18,120 Speaker 1: this episode for forward Thinking. It was in the form 197 00:13:18,400 --> 00:13:21,520 Speaker 1: of analyzing some text and telling you whether or not 198 00:13:22,040 --> 00:13:26,120 Speaker 1: that text would come across as agreeable or argumentative, or 199 00:13:26,200 --> 00:13:31,439 Speaker 1: positive or negative, and it would assign tone to those pieces. 200 00:13:32,040 --> 00:13:35,040 Speaker 1: I'll explain more about how it did and what it 201 00:13:35,120 --> 00:13:37,560 Speaker 1: did in just a minute, but first let's take a 202 00:13:37,640 --> 00:13:48,360 Speaker 1: quick break to thank our sponsor. So how did this 203 00:13:48,440 --> 00:13:53,920 Speaker 1: tone analyzer work. It would search for cues in any 204 00:13:54,080 --> 00:13:59,480 Speaker 1: written text, social cues, written cues, emotional cues in order 205 00:13:59,520 --> 00:14:02,760 Speaker 1: to determine in the overall tone of a piece, which 206 00:14:02,800 --> 00:14:07,640 Speaker 1: actually meant that The analyzer would tag individual words within 207 00:14:07,960 --> 00:14:13,160 Speaker 1: a text, words that it recognized and had already pre 208 00:14:13,280 --> 00:14:17,319 Speaker 1: labeled as falling into various categories. So words that might 209 00:14:17,360 --> 00:14:23,880 Speaker 1: have a positive meaning like happy, glad, joy, things like that. 210 00:14:23,880 --> 00:14:27,480 Speaker 1: Those would get tagged as cheerful. But then it would 211 00:14:27,480 --> 00:14:31,040 Speaker 1: then assign all the individual words tags and then tally 212 00:14:31,120 --> 00:14:33,680 Speaker 1: everything up. So let's say you've got a bunch of 213 00:14:33,680 --> 00:14:39,000 Speaker 1: sentences and it starts individually labeling certain words as being 214 00:14:39,120 --> 00:14:44,240 Speaker 1: cheerful or sad or angry or helpful, and then it 215 00:14:44,280 --> 00:14:46,680 Speaker 1: adds it all up and then would give you a percentage. 216 00:14:47,120 --> 00:14:52,880 Speaker 1: So a message might be agreeable or thirty conscientious, you 217 00:14:52,880 --> 00:14:55,760 Speaker 1: would actually get multiples of these, and that would just 218 00:14:55,800 --> 00:14:59,600 Speaker 1: really indicate the density of those types of words within 219 00:14:59,640 --> 00:15:04,240 Speaker 1: the mess itage itself. Now, in an ideal world, if 220 00:15:04,320 --> 00:15:08,960 Speaker 1: language were very simple to understand and interpret by machines, 221 00:15:09,480 --> 00:15:12,960 Speaker 1: this would help you gauge how people would respond to 222 00:15:13,080 --> 00:15:17,360 Speaker 1: your work. Right, So, you could write a message. Before 223 00:15:17,400 --> 00:15:20,400 Speaker 1: you send it, you put it through the tone analyzer 224 00:15:20,800 --> 00:15:25,000 Speaker 1: and it tells you what sort of a tone you 225 00:15:25,040 --> 00:15:28,360 Speaker 1: are setting. So if you wanted to create a business letter, 226 00:15:28,960 --> 00:15:30,840 Speaker 1: you could send it through this tone analyzer, and if 227 00:15:30,840 --> 00:15:33,760 Speaker 1: it came back as saying it's coming across as as 228 00:15:33,840 --> 00:15:37,320 Speaker 1: a indecisive, you might want to go back in and 229 00:15:37,480 --> 00:15:40,680 Speaker 1: edit that message so that you can make a more 230 00:15:41,080 --> 00:15:46,640 Speaker 1: straightforward and decisive message and not give the wrong impression 231 00:15:46,720 --> 00:15:50,320 Speaker 1: before you send the message out to your actual human recipient, 232 00:15:50,680 --> 00:15:53,280 Speaker 1: and come up with alternate word choices in order to 233 00:15:53,280 --> 00:15:55,200 Speaker 1: make sure that your message is received the way you 234 00:15:55,240 --> 00:15:58,560 Speaker 1: intended it. And anyone who has communicated over the internet 235 00:15:58,600 --> 00:16:01,280 Speaker 1: can think of ways that this might have been helpful 236 00:16:01,320 --> 00:16:05,400 Speaker 1: in the past, because again, language depends on so many 237 00:16:05,520 --> 00:16:09,800 Speaker 1: different elements to get your meaning across, and when you 238 00:16:09,840 --> 00:16:14,520 Speaker 1: reduce it to the written form, especially the written form online, 239 00:16:14,560 --> 00:16:19,239 Speaker 1: where we tend to be very short with our our communication, 240 00:16:19,400 --> 00:16:22,880 Speaker 1: it comes in very quick bursts, a couple of sentences 241 00:16:22,880 --> 00:16:25,960 Speaker 1: here or there. We lack all that body language, we 242 00:16:26,040 --> 00:16:29,320 Speaker 1: lack that tone. It's very easy to misinterpret. I'm sure 243 00:16:29,360 --> 00:16:32,440 Speaker 1: there's been an example in your life where either you 244 00:16:32,520 --> 00:16:35,080 Speaker 1: got offended from receiving something that was meant in a 245 00:16:35,120 --> 00:16:38,360 Speaker 1: way that was different from the way you you interpreted it, 246 00:16:38,480 --> 00:16:40,920 Speaker 1: or the reverse happened where you sent a message and 247 00:16:41,000 --> 00:16:45,320 Speaker 1: somebody had a reaction you did not anticipate because they 248 00:16:45,360 --> 00:16:48,240 Speaker 1: could not tell what tone you were using just from 249 00:16:48,280 --> 00:16:51,960 Speaker 1: the words you were using. Machines have that same problem. 250 00:16:52,200 --> 00:16:55,760 Speaker 1: In the future, an analyzer like this tone analyzer, it 251 00:16:55,760 --> 00:17:00,280 Speaker 1: could be incorporated into word processors or email sir verse, 252 00:17:00,360 --> 00:17:03,920 Speaker 1: or email services, I should say, or social media platforms. 253 00:17:04,240 --> 00:17:06,879 Speaker 1: So you start typing in your message, and before you 254 00:17:06,960 --> 00:17:11,159 Speaker 1: hit published or post or send, you could analyze that text. 255 00:17:11,680 --> 00:17:13,560 Speaker 1: It could tell you what the tone is, and then 256 00:17:13,600 --> 00:17:16,440 Speaker 1: you could say, oh, no, that's gonna come across totally 257 00:17:16,600 --> 00:17:18,840 Speaker 1: the wrong way, and you could actually fix it before 258 00:17:18,920 --> 00:17:21,000 Speaker 1: you posted it or sent it, and then you wouldn't 259 00:17:21,040 --> 00:17:24,680 Speaker 1: have that awkward decision of whether or not to edit something, or, 260 00:17:24,720 --> 00:17:27,639 Speaker 1: in the case of Twitter, which continues to refuse to 261 00:17:27,680 --> 00:17:30,919 Speaker 1: allow you to edit tweets, to delete a tweet. I 262 00:17:31,000 --> 00:17:33,960 Speaker 1: deleted a tweet the other day when I posted a 263 00:17:34,040 --> 00:17:36,679 Speaker 1: link to a news story, and I had done a 264 00:17:36,760 --> 00:17:40,080 Speaker 1: rookie mistake, one that I try to avoid, but I 265 00:17:40,640 --> 00:17:43,800 Speaker 1: did it this pastime, which is that I didn't think 266 00:17:43,840 --> 00:17:46,040 Speaker 1: to look at the date when the news item had 267 00:17:46,080 --> 00:17:49,240 Speaker 1: been published, and had been published a full year earlier, 268 00:17:49,600 --> 00:17:51,919 Speaker 1: so it was not new news, it was old news. 269 00:17:52,440 --> 00:17:55,240 Speaker 1: And uh then deleted the tweet and it wasn't up 270 00:17:55,280 --> 00:17:57,520 Speaker 1: for long, but I still felt dumb about it. It 271 00:17:57,520 --> 00:17:59,239 Speaker 1: would have been nice to have been able to check that. 272 00:17:59,440 --> 00:18:02,119 Speaker 1: Although that's not tone obviously, that's but similar in the 273 00:18:02,840 --> 00:18:06,200 Speaker 1: and the idea that you want to check before you 274 00:18:06,920 --> 00:18:10,240 Speaker 1: end up offending someone, unless you're one of those jerk 275 00:18:10,320 --> 00:18:13,000 Speaker 1: faces that just sets out to offend people, in which case, 276 00:18:14,000 --> 00:18:16,960 Speaker 1: rethink your strategy. There are better things to do. It's 277 00:18:17,080 --> 00:18:19,240 Speaker 1: just as you can make just as big an impact 278 00:18:19,320 --> 00:18:21,960 Speaker 1: being a positive person as you can being a jerk face. 279 00:18:22,320 --> 00:18:23,960 Speaker 1: I know it can seem like it's more work, but 280 00:18:24,000 --> 00:18:27,600 Speaker 1: it's also more rewarding in the long run. Okay, soapbox done. So. 281 00:18:27,960 --> 00:18:31,440 Speaker 1: There is a demo of the tone analyzer that's available online, 282 00:18:32,080 --> 00:18:36,080 Speaker 1: and back when we were recording Forward Thinking, the demo 283 00:18:36,480 --> 00:18:39,240 Speaker 1: worked in a way where it would tell you about 284 00:18:39,280 --> 00:18:42,760 Speaker 1: emotional tone and break it down by percentage. It's a 285 00:18:42,760 --> 00:18:46,199 Speaker 1: little different now, but I want to tell you the 286 00:18:46,920 --> 00:18:50,639 Speaker 1: what words and the results we got in the past 287 00:18:50,760 --> 00:18:53,840 Speaker 1: because they were so much fun. Granted you would get 288 00:18:53,880 --> 00:18:56,520 Speaker 1: a different result now because the tone analyzer has been 289 00:18:56,560 --> 00:19:00,000 Speaker 1: tweaked since we recorded that episode. So when we recorded 290 00:19:00,040 --> 00:19:03,680 Speaker 1: that episode, one of my co hosts decided to put 291 00:19:03,760 --> 00:19:08,560 Speaker 1: a sentence that is somewhat known in literary circles into 292 00:19:08,560 --> 00:19:10,879 Speaker 1: this tone analyzer and find out what it said. And 293 00:19:10,960 --> 00:19:14,879 Speaker 1: the sentence used was it is a truth universally acknowledged 294 00:19:15,080 --> 00:19:17,640 Speaker 1: that a single man in possession of a good fortune 295 00:19:17,960 --> 00:19:21,240 Speaker 1: must be in want of a wife. Now, the analyzer 296 00:19:21,800 --> 00:19:26,560 Speaker 1: said that this emotional tone was cheerful, the social tone 297 00:19:26,680 --> 00:19:31,000 Speaker 1: was seventy six percent open and fifty agreeable, and the 298 00:19:31,080 --> 00:19:35,760 Speaker 1: writing tone was analytical. You can also view the sentence 299 00:19:35,840 --> 00:19:38,520 Speaker 1: in terms of word count as opposed to the weighted 300 00:19:38,600 --> 00:19:41,840 Speaker 1: value of individual words, and using that view, five percent 301 00:19:41,960 --> 00:19:46,440 Speaker 1: of the sentence sentences were in an emotional tone, in 302 00:19:46,480 --> 00:19:49,879 Speaker 1: a social tone, and five percent in a writing tone. Now, 303 00:19:50,280 --> 00:19:54,240 Speaker 1: the analyzer highlights each word according to how it classifies them, 304 00:19:54,680 --> 00:19:58,520 Speaker 1: So emotional words would be highlighted in red or pink 305 00:19:58,600 --> 00:20:01,439 Speaker 1: in that older version of the tone analyzer, social words 306 00:20:01,680 --> 00:20:05,280 Speaker 1: would show up in blue, and writing tones would be 307 00:20:05,359 --> 00:20:07,879 Speaker 1: in green. And you could click on any word and 308 00:20:07,880 --> 00:20:10,720 Speaker 1: the analyzer would offer alternative words that you might want 309 00:20:10,720 --> 00:20:14,159 Speaker 1: to use and classify those words in the tones that 310 00:20:14,320 --> 00:20:16,639 Speaker 1: they are associated with. Such you could shape your message 311 00:20:16,680 --> 00:20:19,439 Speaker 1: to meet the tone you wish to convey. Also, the 312 00:20:19,560 --> 00:20:24,320 Speaker 1: tone analyzer demo used the business letter format as the 313 00:20:24,320 --> 00:20:28,440 Speaker 1: means of comparison, So, in other words, we compared Jane 314 00:20:28,480 --> 00:20:32,320 Speaker 1: Austen to a business letter. Presumably if you were to 315 00:20:32,480 --> 00:20:34,960 Speaker 1: use a full version of the analyzer, not just the 316 00:20:34,960 --> 00:20:37,720 Speaker 1: demo version. You would have other options so you could 317 00:20:38,080 --> 00:20:42,160 Speaker 1: compare it with other models, not just a business letter 318 00:20:42,600 --> 00:20:49,640 Speaker 1: Joe McCormick. He included an excerpt from Dostoyevsky's Notes from Underground. 319 00:20:49,680 --> 00:20:53,639 Speaker 1: That excerpt was, I could not become anything, neither good 320 00:20:53,680 --> 00:20:57,280 Speaker 1: nor bad, neither a scoundrel nor an honest man, neither 321 00:20:57,359 --> 00:21:00,800 Speaker 1: a hero nor an insect. And now I eking out 322 00:21:00,920 --> 00:21:04,760 Speaker 1: my days in my corner, taunting myself with the bitter 323 00:21:04,960 --> 00:21:09,879 Speaker 1: and entirely useless constellation that an intelligent man cannot seriously 324 00:21:09,960 --> 00:21:14,600 Speaker 1: become anything, that only a fool can become something. The 325 00:21:14,640 --> 00:21:19,480 Speaker 1: feedback was that the emotional tone had anger at cheerfulness 326 00:21:19,560 --> 00:21:24,879 Speaker 1: at so happy anger negative at. The social tone was 327 00:21:25,880 --> 00:21:31,080 Speaker 1: agreeable zero percent conscientious, zero percent open. The writing tone 328 00:21:31,119 --> 00:21:36,600 Speaker 1: was analytical, zero percent confident and tentative. Joe would actually 329 00:21:36,720 --> 00:21:39,760 Speaker 1: end up highlighting some of the words to find out 330 00:21:39,920 --> 00:21:42,359 Speaker 1: which words were the ones that ended up giving that 331 00:21:43,600 --> 00:21:47,920 Speaker 1: cheerfulness result. Those four words were a good, honest, hero, 332 00:21:48,200 --> 00:21:55,040 Speaker 1: and intelligent and that kind of are that that's important 333 00:21:55,280 --> 00:21:59,399 Speaker 1: because those words, the way they are used uh in 334 00:21:59,480 --> 00:22:03,680 Speaker 1: that passage are not used in a positive sense. They 335 00:22:03,720 --> 00:22:09,000 Speaker 1: are positive words, but they're meant to show kind of 336 00:22:09,040 --> 00:22:15,280 Speaker 1: a negation there not, and not an assertion. So that 337 00:22:15,359 --> 00:22:18,720 Speaker 1: really highlights a big problem in this tone analyzer, which 338 00:22:18,760 --> 00:22:24,719 Speaker 1: is that it's tagging these words individually without context. So 339 00:22:24,800 --> 00:22:28,680 Speaker 1: if I wrote the phrase I am not glad, it 340 00:22:28,720 --> 00:22:31,520 Speaker 1: would tag the word glad and say that's a cheerful word. 341 00:22:32,200 --> 00:22:35,879 Speaker 1: But I said I am not glad. You if I 342 00:22:35,960 --> 00:22:38,960 Speaker 1: told you I am not glad, you would not think, oh, well, 343 00:22:38,960 --> 00:22:40,919 Speaker 1: that's a cheerful thing to say or a positive thing 344 00:22:40,960 --> 00:22:44,560 Speaker 1: to say. But according to the tone analyzer, it would 345 00:22:44,600 --> 00:22:47,920 Speaker 1: come across as a cheerful statement because it had tagged 346 00:22:47,920 --> 00:22:50,119 Speaker 1: that word as as being cheerful. In the other words 347 00:22:50,359 --> 00:22:53,880 Speaker 1: are not that strong, they don't they don't warrant being 348 00:22:53,880 --> 00:22:58,280 Speaker 1: tagged in a way like that. Now, over time, we 349 00:22:58,359 --> 00:23:01,360 Speaker 1: might have a tone analyzer that can actually take context 350 00:23:01,600 --> 00:23:05,879 Speaker 1: into account, and then you would learn a lot more 351 00:23:05,920 --> 00:23:09,679 Speaker 1: about the actual meaning behind a phrase. It would be 352 00:23:09,720 --> 00:23:12,520 Speaker 1: more than just tone. So if you were trying to 353 00:23:12,520 --> 00:23:18,240 Speaker 1: get across tone by using more complicated and subtle word choice, 354 00:23:18,760 --> 00:23:23,520 Speaker 1: where you're sort of being kind of uh poetic in 355 00:23:23,560 --> 00:23:28,200 Speaker 1: your expression, you're trying to get across a feeling by 356 00:23:28,280 --> 00:23:33,399 Speaker 1: using irony or sarcasm. Then a tone analyzer like this 357 00:23:33,440 --> 00:23:36,040 Speaker 1: would totally miss it because it would just be counting 358 00:23:36,040 --> 00:23:40,280 Speaker 1: the hits and not understanding the usage there the hidden 359 00:23:40,359 --> 00:23:44,520 Speaker 1: meeting the word play. So that is going to be 360 00:23:44,960 --> 00:23:49,880 Speaker 1: a real challenge. So it's kind of another interesting use 361 00:23:49,880 --> 00:23:52,120 Speaker 1: of IBMS Watson. There are a lot of other ones 362 00:23:52,160 --> 00:23:54,600 Speaker 1: that we could talk about, like Chef Watson, which was 363 00:23:54,680 --> 00:23:58,960 Speaker 1: my favorite. Chef Watson would generate new recipes based upon 364 00:23:59,160 --> 00:24:01,600 Speaker 1: ingredients that you would tell it that you had on hand, 365 00:24:02,040 --> 00:24:07,000 Speaker 1: and it wouldn't it wouldn't go and reference old recipes 366 00:24:07,040 --> 00:24:09,800 Speaker 1: and pull one up for you. Instead, it would make 367 00:24:09,840 --> 00:24:13,520 Speaker 1: flavor profiles based upon all the different combinations of food 368 00:24:13,560 --> 00:24:16,280 Speaker 1: that were found in various recipe books and generate a 369 00:24:16,280 --> 00:24:18,879 Speaker 1: brand new recipe for you, right there on the spot. 370 00:24:19,240 --> 00:24:24,000 Speaker 1: And sometimes they were whacka doodle crazy, y'all. So in 371 00:24:24,040 --> 00:24:26,240 Speaker 1: a way you could say that Chef Watson was another 372 00:24:26,640 --> 00:24:29,760 Speaker 1: another way of seeing how IBM S Watson has a 373 00:24:29,800 --> 00:24:33,480 Speaker 1: lot of promise, but it requires a ton of work 374 00:24:34,000 --> 00:24:37,600 Speaker 1: on the app level in order to leverage it and 375 00:24:37,640 --> 00:24:40,440 Speaker 1: make actual practical use out of it. I have more 376 00:24:40,480 --> 00:24:45,280 Speaker 1: to say about computers detecting sarcasm. But first let's take 377 00:24:45,960 --> 00:24:58,520 Speaker 1: a quick word from our sponsor. So back in twent 378 00:24:59,600 --> 00:25:03,240 Speaker 1: there were some researchers at the Hebrew University in Israel 379 00:25:03,359 --> 00:25:08,760 Speaker 1: who designed a system called the Semi Supervised Algorithm for 380 00:25:08,800 --> 00:25:15,639 Speaker 1: Sarcasm Identification or SAZI, and they used SAZI to analyze 381 00:25:15,640 --> 00:25:20,520 Speaker 1: collections of nearly six million tweets and also around sixty 382 00:25:20,600 --> 00:25:25,680 Speaker 1: six thousand product reviews from Amazon. They wanted to find 383 00:25:26,480 --> 00:25:31,160 Speaker 1: rich treasure troves of sarcasm that turns out reviews and 384 00:25:31,200 --> 00:25:37,119 Speaker 1: tweets they fit the bill sarcasm is. Really it's typically 385 00:25:37,200 --> 00:25:40,960 Speaker 1: conveyed in in some vocal tone right and nonverbal cues. 386 00:25:41,760 --> 00:25:45,840 Speaker 1: So you have to first go someplace where sarcasm is 387 00:25:45,840 --> 00:25:49,240 Speaker 1: is rampant in text form to be able to really 388 00:25:49,400 --> 00:25:54,280 Speaker 1: fine tune how you can identify sarcasm versus something that's 389 00:25:54,320 --> 00:25:57,400 Speaker 1: meant exactly the way it's written on the surface level. 390 00:25:57,760 --> 00:26:03,120 Speaker 1: So they started to map out the various features that 391 00:26:03,200 --> 00:26:07,520 Speaker 1: were common in sarcastic comments online. So they were looking 392 00:26:07,520 --> 00:26:11,520 Speaker 1: for things like hyperbolic words and if you're using a 393 00:26:11,520 --> 00:26:15,440 Speaker 1: lot of exaggeration, that could be a key. Excessive punctuation 394 00:26:15,760 --> 00:26:19,040 Speaker 1: was another one, especially ellipses, which I tend to use 395 00:26:19,160 --> 00:26:21,480 Speaker 1: a lot, though I don't know if I use it 396 00:26:21,520 --> 00:26:24,680 Speaker 1: so much for sarcasm as I do for just timing purposes. 397 00:26:24,720 --> 00:26:27,399 Speaker 1: To indicate this is the beat I would take if 398 00:26:27,400 --> 00:26:30,159 Speaker 1: I were saying this out loud. I guess that's just 399 00:26:30,240 --> 00:26:34,560 Speaker 1: as irritating, though, also how straightforward is the Senate structure? 400 00:26:35,040 --> 00:26:37,600 Speaker 1: And they gave it examples of sarcasm. They fed it 401 00:26:37,680 --> 00:26:43,919 Speaker 1: tweets that were tagged hashtag sarcasm, so that the machine 402 00:26:43,960 --> 00:26:47,600 Speaker 1: quote unquote knew that that was already a sarcastic tweet 403 00:26:47,840 --> 00:26:50,919 Speaker 1: and could start to analyze it and build out a 404 00:26:51,040 --> 00:26:53,240 Speaker 1: model for what sarcasm is. They also fed at a 405 00:26:53,320 --> 00:26:57,080 Speaker 1: bunch of one star Amazon reviews that had been judged 406 00:26:57,160 --> 00:27:01,480 Speaker 1: to be sarcastic by a panel consisting of fifteen human beings, 407 00:27:02,040 --> 00:27:06,880 Speaker 1: and the system was told it had to rate sentences 408 00:27:06,920 --> 00:27:10,440 Speaker 1: on a scale of one to five, One being not sarcastic. 409 00:27:10,880 --> 00:27:16,040 Speaker 1: They mean exactly what the Senate says, five being holy cow, 410 00:27:16,200 --> 00:27:20,440 Speaker 1: this person should write for the Onion, this is incredibly sarcastic. 411 00:27:21,000 --> 00:27:27,800 Speaker 1: SAZI could identify sarcastic Amazon reviews with precision, not bad, 412 00:27:28,840 --> 00:27:31,440 Speaker 1: but when it came to Twitter, it did even better, 413 00:27:32,200 --> 00:27:36,159 Speaker 1: I think, probably because there had to be very short 414 00:27:36,200 --> 00:27:39,280 Speaker 1: messages on Twitter. This was before Twitter had even expanded 415 00:27:39,280 --> 00:27:42,560 Speaker 1: to characters, so it's still back in the one character days. 416 00:27:43,080 --> 00:27:47,760 Speaker 1: The precision rate for SAZI for Twitter was so it 417 00:27:47,840 --> 00:27:52,560 Speaker 1: was really good at detecting straightforward sarcasm, the kind that 418 00:27:52,600 --> 00:27:55,000 Speaker 1: a lot of people on Twitter use, because you have 419 00:27:55,160 --> 00:27:57,240 Speaker 1: limited space so you can't really set it up in 420 00:27:57,320 --> 00:28:01,720 Speaker 1: a more complex way, but it was all so uh 421 00:28:02,080 --> 00:28:08,199 Speaker 1: more prone to judging things as false negative evaluations rather 422 00:28:08,240 --> 00:28:10,960 Speaker 1: than false positives. In other words, it was more likely 423 00:28:11,600 --> 00:28:15,600 Speaker 1: to look at a negative sarcastic message and say that's 424 00:28:15,600 --> 00:28:18,440 Speaker 1: not sarcastic than it was to look at a straightforward 425 00:28:18,440 --> 00:28:21,960 Speaker 1: message and say, no, that is sarcastic. So that was 426 00:28:22,040 --> 00:28:27,040 Speaker 1: kind of interesting. Back to Watson. Another use of Watson 427 00:28:27,480 --> 00:28:31,520 Speaker 1: came out of the Milk and Institute Global Conference at 428 00:28:31,720 --> 00:28:35,720 Speaker 1: IBM showed off some research that it had been working 429 00:28:35,840 --> 00:28:40,520 Speaker 1: on internally, and it was calling this research debating Technologies. 430 00:28:41,280 --> 00:28:44,600 Speaker 1: This was a project in which IBM was trying to 431 00:28:44,640 --> 00:28:48,440 Speaker 1: see if they could feed a computer raw information, have 432 00:28:48,640 --> 00:28:53,640 Speaker 1: the computer synthesize the information, understand that information, at least 433 00:28:53,640 --> 00:29:00,840 Speaker 1: on a computational level and then create a a debating 434 00:29:00,880 --> 00:29:05,000 Speaker 1: strategy for both pros and cons based on that information. 435 00:29:05,400 --> 00:29:09,080 Speaker 1: So it would take a huge amount of content like 436 00:29:09,720 --> 00:29:13,280 Speaker 1: all of Wikipedia, for example, and then on any given 437 00:29:13,280 --> 00:29:15,920 Speaker 1: subject that would be covered in Wikipedia, it would be 438 00:29:15,960 --> 00:29:19,800 Speaker 1: asked form an argument that is in favor of or 439 00:29:19,960 --> 00:29:25,000 Speaker 1: is against a concept, whatever that concept might be. John 440 00:29:25,120 --> 00:29:27,560 Speaker 1: Kelly of IBM showed off in a demo how the 441 00:29:27,560 --> 00:29:31,080 Speaker 1: tool could be used to predict pro or con arguments 442 00:29:31,120 --> 00:29:35,360 Speaker 1: about a subject based on a body of information. So 443 00:29:36,400 --> 00:29:40,360 Speaker 1: you might be able to use this technology in order 444 00:29:40,400 --> 00:29:47,000 Speaker 1: to anticipate what an opposing person might say on any 445 00:29:47,040 --> 00:29:49,360 Speaker 1: given subject. Let's say that you are getting ready to 446 00:29:49,440 --> 00:29:55,200 Speaker 1: debate a topic. You might feed that information to a 447 00:29:55,280 --> 00:29:58,480 Speaker 1: computer system using this Watson platform. You might feed in 448 00:29:58,560 --> 00:30:02,400 Speaker 1: a ton of information, and then you might say, who 449 00:30:03,760 --> 00:30:08,000 Speaker 1: imagine someone who is against this particular topic, whatever it 450 00:30:08,080 --> 00:30:12,560 Speaker 1: might be. Uh. Let's say it's it's it's renewable energy 451 00:30:12,960 --> 00:30:17,040 Speaker 1: and the uh the efficiency of solar panels, whether or 452 00:30:17,040 --> 00:30:20,000 Speaker 1: not it makes sense to invest in solar panels. Let's 453 00:30:20,000 --> 00:30:22,480 Speaker 1: say that your stance is that you have to argue 454 00:30:22,640 --> 00:30:26,200 Speaker 1: for solar panels. You might say, what would someone who 455 00:30:26,200 --> 00:30:31,040 Speaker 1: wants to argue against solar panels, say, and then Watson 456 00:30:31,120 --> 00:30:36,160 Speaker 1: would analyze this information and return to you what it 457 00:30:36,280 --> 00:30:40,480 Speaker 1: thinks would be an argument someone would use to support 458 00:30:40,560 --> 00:30:45,040 Speaker 1: that that stance, and then you could prepare for that, 459 00:30:45,640 --> 00:30:47,640 Speaker 1: which would be an incredible tool. I mean, you could 460 00:30:47,640 --> 00:30:50,000 Speaker 1: think of this as for political debates. It would be amazing. 461 00:30:50,200 --> 00:30:53,000 Speaker 1: You could think of how you might want to prepare 462 00:30:53,320 --> 00:30:56,480 Speaker 1: so that you can argue intelligently against an opponent, and 463 00:30:56,480 --> 00:30:58,920 Speaker 1: you can already anticipate what that opponent is going to 464 00:30:58,960 --> 00:31:01,959 Speaker 1: say because you oh their general stance on a topic, 465 00:31:02,240 --> 00:31:04,760 Speaker 1: but you might not know what tactics they might use 466 00:31:04,840 --> 00:31:08,760 Speaker 1: to support that stance. Maybe politics isn't a great choice 467 00:31:08,800 --> 00:31:11,440 Speaker 1: because that's not always in the realm of rationality. That 468 00:31:11,480 --> 00:31:17,840 Speaker 1: often falls into a call toward emotional response rather than 469 00:31:17,960 --> 00:31:22,640 Speaker 1: rational response. That's more of a a commentary on politics 470 00:31:22,640 --> 00:31:25,440 Speaker 1: in general, regardless of what side you might be on, 471 00:31:25,680 --> 00:31:29,240 Speaker 1: all sides do this anyway. He actually showed at this 472 00:31:29,320 --> 00:31:32,680 Speaker 1: demo a different example. He said, what if you were 473 00:31:32,720 --> 00:31:35,800 Speaker 1: to take the sale of violent video games to minors 474 00:31:35,960 --> 00:31:40,280 Speaker 1: should be banned, that's the topic, and that the computer 475 00:31:40,320 --> 00:31:43,040 Speaker 1: would then go through all the information and had access 476 00:31:43,080 --> 00:31:46,720 Speaker 1: to it would end up sorting out all the parts 477 00:31:46,760 --> 00:31:49,840 Speaker 1: that were relevant to the discussion, so it just put 478 00:31:49,880 --> 00:31:52,840 Speaker 1: those aside and that would become the core of the 479 00:31:52,920 --> 00:31:55,960 Speaker 1: data it would reference. I would then go through and 480 00:31:56,040 --> 00:32:00,760 Speaker 1: identify basic statements is either being a pro stance of 481 00:32:01,880 --> 00:32:07,080 Speaker 1: banning violent video games to miners or a constance for 482 00:32:07,160 --> 00:32:09,680 Speaker 1: that saying no, we should be able to sell violent 483 00:32:09,760 --> 00:32:14,520 Speaker 1: video games to miners. The tools scanned four million articles. 484 00:32:14,560 --> 00:32:17,280 Speaker 1: It returned the top ten articles that were determined to 485 00:32:17,320 --> 00:32:21,200 Speaker 1: be the most relevant to that particular debate, and it 486 00:32:21,320 --> 00:32:26,280 Speaker 1: scanned approximately three thousand sentences, come from from top to bottom, 487 00:32:26,440 --> 00:32:31,000 Speaker 1: and it then identified sentences that contained candidate claims that 488 00:32:31,320 --> 00:32:34,560 Speaker 1: would be statements that would either be interpreted as being 489 00:32:34,600 --> 00:32:37,920 Speaker 1: pro or con for the stance. Then it identified the 490 00:32:37,920 --> 00:32:41,000 Speaker 1: parameters of those claims. Then it assessed the claims for 491 00:32:41,120 --> 00:32:44,400 Speaker 1: the pro and con polarity, then constructed a sample pro 492 00:32:44,640 --> 00:32:47,840 Speaker 1: or con statement. And the statements in the demo were 493 00:32:48,240 --> 00:32:51,360 Speaker 1: kind of interesting. And since the computer is constructing arguments 494 00:32:51,440 --> 00:32:55,880 Speaker 1: based upon what people have already written, it would reflect 495 00:32:55,880 --> 00:32:58,960 Speaker 1: a lot of vague statements that aren't a firm stance. So, 496 00:32:59,000 --> 00:33:01,080 Speaker 1: in other words, like it it and take a bunch 497 00:33:01,320 --> 00:33:05,560 Speaker 1: of stuff that was written that itself did not take 498 00:33:05,920 --> 00:33:09,440 Speaker 1: either a pro or constance and then transform that magically 499 00:33:09,640 --> 00:33:13,200 Speaker 1: into the perfect pro stance or the perfect constance. Uh. 500 00:33:13,280 --> 00:33:16,360 Speaker 1: It's dependent upon the words that human beings have already written, 501 00:33:16,800 --> 00:33:19,600 Speaker 1: So it could not magically come up with a killer 502 00:33:19,960 --> 00:33:22,760 Speaker 1: argument if the data that had been written about this 503 00:33:22,920 --> 00:33:27,280 Speaker 1: subject didn't come down on a firm stance one way 504 00:33:27,360 --> 00:33:32,560 Speaker 1: or the other. Um. The point of the demonstration wasn't 505 00:33:32,600 --> 00:33:36,800 Speaker 1: to create a tool that could either troll people or 506 00:33:36,960 --> 00:33:39,880 Speaker 1: counter trolls. It was to show that a computer could 507 00:33:39,960 --> 00:33:43,120 Speaker 1: be useful to aid in the reasoning process when you're 508 00:33:43,200 --> 00:33:46,520 Speaker 1: making a critical decision. Again, to go back to that 509 00:33:46,600 --> 00:33:49,600 Speaker 1: medical example, it could be used to help a doctor 510 00:33:50,120 --> 00:33:54,320 Speaker 1: determine which diagnosis is the most likely to be accurate 511 00:33:54,360 --> 00:33:58,240 Speaker 1: for a patient, what what course of treatment might be 512 00:33:58,320 --> 00:34:03,160 Speaker 1: the most helpful for that patient, and thus it could 513 00:34:03,240 --> 00:34:07,480 Speaker 1: have real practical use outside of this more esoteric, interesting 514 00:34:08,239 --> 00:34:13,160 Speaker 1: UH debate. Us. Now, will we see computers in the 515 00:34:13,280 --> 00:34:17,759 Speaker 1: future able to detect sarcasm just as easily as your 516 00:34:17,760 --> 00:34:23,440 Speaker 1: typical human being can when given the right circumstances. And 517 00:34:23,480 --> 00:34:27,040 Speaker 1: I use the word typical reluctantly, but you get what 518 00:34:27,080 --> 00:34:30,439 Speaker 1: I mean, I don't know. It's gonna take some time. 519 00:34:30,719 --> 00:34:32,960 Speaker 1: It takes an awful lot of processing power too. You 520 00:34:32,960 --> 00:34:37,400 Speaker 1: have to remember that for these neural networks systems, the 521 00:34:37,400 --> 00:34:40,879 Speaker 1: ones that are running these these various platforms and programs 522 00:34:40,880 --> 00:34:44,600 Speaker 1: and strategies, they take up a lot of processing power. 523 00:34:45,840 --> 00:34:52,120 Speaker 1: Because our brains have billion neurons in them, so we 524 00:34:52,200 --> 00:34:56,560 Speaker 1: have a very sophisticated supercomputer sitting in our heads. Moreover, 525 00:34:56,960 --> 00:35:01,160 Speaker 1: our brains are insanely energy efficient. They require about the 526 00:35:01,200 --> 00:35:05,239 Speaker 1: equivalent of twenty watts of power. A supercomputer needs a 527 00:35:05,320 --> 00:35:09,680 Speaker 1: lot more power than that. So while we're seeing advances 528 00:35:09,719 --> 00:35:13,720 Speaker 1: in this, it requires so much processing power, so much energy. 529 00:35:14,080 --> 00:35:19,319 Speaker 1: It is not a practical approach to most forms of computing, 530 00:35:19,800 --> 00:35:23,400 Speaker 1: at least from a consumer standpoint. You might see a 531 00:35:23,440 --> 00:35:25,759 Speaker 1: future where the sort of stuff is all in the 532 00:35:25,800 --> 00:35:29,720 Speaker 1: cloud and then we can access it through an app 533 00:35:29,840 --> 00:35:32,239 Speaker 1: or a program or whatever. That way, you don't have 534 00:35:32,320 --> 00:35:35,160 Speaker 1: to have a supercomputer sitting on your desk in order 535 00:35:35,160 --> 00:35:38,239 Speaker 1: to tap into those uh, those capabilities, but you have 536 00:35:38,280 --> 00:35:41,239 Speaker 1: to have an Internet connection, which most of us these 537 00:35:41,320 --> 00:35:44,200 Speaker 1: days tend to have fairly frequently. I mean, there are 538 00:35:44,200 --> 00:35:46,200 Speaker 1: a lot of people out there who at this point 539 00:35:46,520 --> 00:35:49,680 Speaker 1: have had a persistent Internet connection for pretty much their 540 00:35:49,680 --> 00:35:53,440 Speaker 1: whole lives, which blows my mind. But that's the kind 541 00:35:53,480 --> 00:35:55,480 Speaker 1: of world we'd have to live in in order to 542 00:35:55,560 --> 00:35:58,640 Speaker 1: really take advantage of this, at least in the near term. 543 00:35:58,680 --> 00:36:00,560 Speaker 1: I don't know if we're are going to see a 544 00:36:00,600 --> 00:36:04,480 Speaker 1: computer that can analyze, say, an article from the Onion 545 00:36:05,280 --> 00:36:09,200 Speaker 1: and not only point out that it's being sarcastic or ironic, 546 00:36:09,320 --> 00:36:11,920 Speaker 1: but also point out why it's funny. I think at 547 00:36:11,960 --> 00:36:14,759 Speaker 1: one point, when you start analyzing comedy, that gets to 548 00:36:14,840 --> 00:36:18,400 Speaker 1: be a level where nothing is ever funny ever again. 549 00:36:18,920 --> 00:36:23,440 Speaker 1: But it is a really interesting problem. So that's whether 550 00:36:23,520 --> 00:36:26,319 Speaker 1: that's that's this look back on if AI is ever 551 00:36:26,360 --> 00:36:28,920 Speaker 1: going to understand sarcasm. I'm curious to hear what you 552 00:36:28,960 --> 00:36:34,719 Speaker 1: guys think. Do you think we're closer than I am suggesting? Uh? Maybe, well, 553 00:36:34,760 --> 00:36:36,480 Speaker 1: I mean, we're definitely closer than we were when we 554 00:36:36,520 --> 00:36:38,320 Speaker 1: did this episode on Forward Thinking, because that was a 555 00:36:38,320 --> 00:36:42,200 Speaker 1: few years ago. But I don't know that we're you know, 556 00:36:42,320 --> 00:36:46,160 Speaker 1: significantly closer. It's a it's a real tough problem. Or 557 00:36:46,200 --> 00:36:48,279 Speaker 1: do you think that sarcasm is one of those things 558 00:36:48,360 --> 00:36:50,640 Speaker 1: that's just innately human and machines are never really going 559 00:36:50,680 --> 00:36:53,520 Speaker 1: to be able to handle it. We've got a lot 560 00:36:53,600 --> 00:36:56,160 Speaker 1: of programs out there that appear to be sarcastic, but 561 00:36:56,200 --> 00:37:01,080 Speaker 1: that's because they're they're acting on preprogrammed respond says two 562 00:37:01,080 --> 00:37:03,480 Speaker 1: things that we ask them. It's not exactly the same. 563 00:37:03,480 --> 00:37:06,080 Speaker 1: It's kind of cheating, but I'm curious to hear what 564 00:37:06,120 --> 00:37:09,080 Speaker 1: you guys think. Also, make sure you go to our 565 00:37:09,160 --> 00:37:13,359 Speaker 1: brand new website for tech stuff. That's tech Stuff Podcast 566 00:37:13,560 --> 00:37:16,160 Speaker 1: dot com. That's where you're going to find all the 567 00:37:16,239 --> 00:37:18,480 Speaker 1: links to all sorts of stuff like how to contact 568 00:37:18,520 --> 00:37:21,640 Speaker 1: me in case you're wondering the best ways through email, 569 00:37:21,680 --> 00:37:23,759 Speaker 1: It's tech Stuff at how stuff Works dot com, or 570 00:37:23,800 --> 00:37:25,880 Speaker 1: through Facebook or Twitter that's Tech Stuff hs W. But 571 00:37:25,920 --> 00:37:28,320 Speaker 1: all that information is also on the website, as is 572 00:37:28,360 --> 00:37:31,040 Speaker 1: a link to our store at t Public. Remember every 573 00:37:31,080 --> 00:37:34,280 Speaker 1: single purchase you make at that store helps out the show. 574 00:37:34,880 --> 00:37:38,080 Speaker 1: Don't forget to follow us on Instagram and I'll talk 575 00:37:38,120 --> 00:37:46,640 Speaker 1: to you again really soon for more on this and 576 00:37:46,680 --> 00:37:59,359 Speaker 1: thousands of other topics. Because it how Stuff Works dot com.