Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Brought to you by Toyota. Let's go places. Welcome to
Forward Thinking. Hey there, and welcome to Forward Thinking, the
podcast that looks at the future and says no dark
sarcasm in the classroom. I'm Jonathan Strickland and I'm Joe McCormick. So, guys,
(00:23):
I thought today we could talk a little bit more
about how, you know, computers just don't get me, man,
they just don't understand. They're like parents. You know, parents
never understand the importance of the flavor of apple Jack's.
That's true, among many other things. True. In fact, they
also don't understand our music, right, they don't. They don't
(00:46):
get my clothes, my style. Uh. Now, we're we're talking
about how computers, you know, trying to teach computers how
to understand what we mean when we say things. And
we've talked about this before. We've talked about computers in
natural language and how that's a really tough problem getting
computers to understand the way we humans communicating naturally yea,
(01:09):
and especially trying to get them to mimic it or
or feed it back to us and participate in a
conversation that's convincingly human. That's the standard problem of the
touring test. Yeah, even if it's not convincingly human, at
least useful, right like we you know, I would settle
for a computer that would understand what I mean when
I say things a certain way and give me the
(01:31):
stuff that I'm requesting, even if it can't hold a
conversation with me. Right right, We don't always need Siri
to to respond back if if she could just find
what we're bloody looking for. Cortana would be another great
example when those ten has just been released the day
we're recording this, and I'm seeing a lot of reports
about how Cortana is almost but not quite awesome. You know.
(01:53):
One of the things that's interesting to me about Siri
is that Siri is playful Siri. Siri will sometimes be
koy and sometimes like play along when you're when you're
being I don't know, suggestive or weird at her. I
don't say this from personal experience. I've read about this,
but I get the impression that that is not original
(02:17):
like synthetic behavior, that it has been assembled by series
logical engines. That is more of a hard coded behavior.
I think that's more of their sort of jokes put
in directly by the programmer. The programmers anticipate the kind
of like they just say, Well, if I found out
that this audio personal assistant could answer questions that I
(02:38):
ask of it, what are some of the ridiculous questions
I might ask? And they probably made a list and
and that's probably the basis for a lot of those
funny responses, things like where can I hide a body? Yeah? Yeah,
but but that's all coded by your word choice. It's
not like it's not like, you know, Siri is gonna
listen to you say something like, h Siri, where can
I find another pizza place? Because I'm so interested in
(03:00):
pizza today? Like she She's not going to come back
with with with like, well, maybe you should try some
chicken nuggets this time, fatty, Like I mean, like I
was thinking, it's Taco Tuesday, chicken nuggets. The cure to
facts that would be a revolutionary episode forward thinking. However,
we wanted to talk today about I learned it from Laura.
(03:24):
We wanted to talk today more about how computers might
soon or at least down the road, learned to recognize
not just what we say, but what we mean when
we say it. Right, So could Siri participate in a
sarcastic sort of caustic back and forth humorous exchange with
(03:45):
you without that being hard coded by the programmers, could
could Siri detected and even and even synthesize it herself. Right,
So to start this, we thought it might be useful
to kind of give a quick overview of why there's
this us connect between natural language and what computers understand
and basically computers machines work in uh in machine code
(04:07):
or assembly language. Assembly language is generally defined as a
very low level programming language, sometimes a one to one
correspondence with the architecture of the actual machine you are
using it on. So the thing about this, the defining characteristic,
is that computers understand it and humans have real trouble
with it because it's it's so far removed from any
(04:29):
sort of language that we use. It would be like
looking at a just enormous page, just a block of
texts that are all zeros and ones, and trying to
make any sense of it. Um. I assume that most
human beings, the vast majority, would be unable to do so.
So we humans communicate natural language. That's what I'm doing
(04:50):
right now. And natural language isn't just agree and agreed
upon syntax and grammar. It also has all these other
elements to it, implications where you know, you don't say
something outright, but it is implied by the way you
say it, or the tone of how you deliver something,
including sarcasms. Sure, even the gestures that you're making or
the facial expression. Yeah. I think we talked pretty recently
(05:13):
on this program about how one major difference is that
humans learn language by induction rather than by you know,
we don't have an explicit list of rules and uh
and definitions, Like we don't learn how to speak our
language by reading a dictionary in a grammar book. You
just sort of gradually into it the rules based on
(05:33):
experiencing them. Yeah. Yeah, and we free associate a lot
of stuff with a lot of other stuff. That's true. Yeah.
So one thing that we have to do is find
a way to bridge that gap between the way we
humans communicate and the way machines understand commands. And we
do that through programming languages and compilers. A programming language,
depending upon whether you call it low level or high level,
(05:54):
is something that may be easier for a human to grasp.
In general, if we say it's a high level programming language.
That means it's more like the kind of languages we use,
although if you're unfamiliar with programming language, it still looks
like gibberish exactly. Yeah, this is a funny fact about
programming languages. People often think of them as something that
helps the computer understand. That's not what it is. It's
(06:17):
something that helps the human understanding its programming language is
a tool for you, right, it's so that the human
can say, uh, this is this is the input I'm
going to give you, this is the desired output, make
it go. And so on the machine side they have
there are compilers and compilers. The job of a compiler
is to essentially translate, to switch from that programming language
(06:40):
into machine language. So it's this combination that allows machines
to understand what we want them to do. But more
recent work, really this has been going on for decades,
but truly in the last decade we've seen a lot
of work done in trying to come up with ways
where machines can deal with natural language and the and
if it is obvious, because it removes a barrier between use,
(07:04):
you know, a human trying to use a machine, and
the machine doing what what the human wants it to do. So,
for a very simple example, if you go on your
computer and you wanted to find a particular piece of information,
you could just type something in and it would understand
what you meant by that, so it wouldn't return a
page of search results that are in some way related
(07:25):
to what you wanted. It would return exactly what you wanted. Uh.
So that that's a simple example. There are lots of
other examples of this, but we're still just at the
very early stages of getting a good grasp on natural language.
We're getting better at computers parsing sentences to at least
make an educated guess as to what we want, but
(07:48):
we're not all the way there yet. But we wanted
to talk about some interesting tools that are getting us closer,
and one was one that was reported on in July
two thousands of team that I really wanted to talk about,
which was IBMS tone Analyzer. Yeah. I had some fun
playing with this earlier. Today. We can talk about the
results I came up with later, but I think a
(08:11):
common theme in our analysis of this will be that
it's interesting, but maybe more for the reasons that it
fails than for the reasons it succeeds. Yeah, I would
agree with that. It's kind of a very advanced version
of Clippy almost in that way I'm like, oh, it's
really cute. The way that you have no idea what's
going on. I almost think of it as as like
(08:31):
and of evolution of just word count. Yeah, like it's
still counting words, it's just now classifying what word is,
what like or what general categories words fall into. But
it's part of the Watson development cloud. Uh. And that's
that if you are familiar with IBM S Watson, you know,
you know that that refers to the machine that played
(08:52):
on Jeopardy and beat to Jeopardy Champions. Yeah, which makes
it hilarious that we're using words like cute to describe
something is so ceated with Watson yea one of the
most amazing computers ever put together. It's really a function
of how difficult the task is. Yeah, more so than
than how dumb this that's not that's not what I'm saying,
(09:13):
No offense, wats no, no, totally that I'm on the
same page. Because Watson was obviously amazing. It was able
to take Jeopardy clues which are in the form of
an answer and come up with the appropriate question more
frequently than not. And we've talked a little bit about
how it did this, and that it was able to
come up with potential answers or rather questions if you prefer,
(09:35):
and then judge how probable that particular response was to
being the correct one, and if it met a certain threshold,
Watson would buzz in and offer that up as the
response to the clue. Yeah. And so what made Watson
interesting was different than what makes most Jeopardy contestants interesting.
It was because most Jeopardy contestants are competing based on
(09:59):
their essentially their volume of trivia that they have contained
in their brain. They're not having trouble discerning what the
clue is trying to say, right, yeah, yeah, like like
they get the puns in it, and and they get
the sarcasm or whatever it is. I guess Alex Trebeck
isn't sarcastic very often. But you know, references as well,
(10:21):
they cultural references contest. Yeah, right, So for them what's
impressive is, Wow, I can't believe that lady knew all
that stuff about ancient Rome and about you know, records
and about video games whatever they talk about. Yeah, but
you wouldn't be impressed at all that she, you know,
makes sense of what the clue is saying. And that's
(10:43):
the deal with Watson. What it's not interesting that a
computer has all that knowledge because it can just you know,
have terabytes of memory that it runs through. What's interesting
is that it knew what the question was referring to, right,
and that that was one of those things that really
opened up a lot of eyes and said, wow, this
is exciting. A computer is, at least on some level
(11:05):
understanding what it's supposed to be looking for. Whether that
you know, that understanding doesn't go on to the same
level as it would with a human, but it's still
really impressive because generally, you would have a computer and
you'd give it some sort of word play response, and
chances are it would not come up with an appropriate answer.
It would either not answer at all because it wouldn't
(11:26):
hit that threshold of certainty, or it would give them
an incorrect answer. And you know, once and once or
twice Watson gave funny wrong answers. It's not like it
was infallible, but it was. It worked more than it
didn't work and in fact beat the two champions, so
pretty exciting. Well, the Watson Developer Cloud is UH sort
(11:47):
of it's it's kind of an umbrella. It's a it's
an application program interface also known as an a p I,
and developers can use it to leverage this cognitive computing approach,
this language wrecking mission approach natural language recognition UH, and
then leverage that into other applications. So Watson itself has
(12:08):
gone on to do lots of work in other fields,
and most notably medicine, being able to help doctors end
up looking at various means of treatment, even personalizing treatment
for patients with certain illnesses or conditions. But this was
something where you know, if you were a developer and
(12:28):
you wanted to try and leverage this technology that had
already been built, you could then build your own thing
and make something really special. And one of those things
was the Tone Analyzer UM. So it it searches for social,
written and emotional cues and a body of text in
order to analyze the tone of the overall text and
(12:49):
to tell you how is that coming across in an
analytical way, Like you know, I I have tested the
block that you have written and it comes across as
seventy three percent cheerful, which sounds really weird, but that's
essentially what you're getting. Yeah, yeah, And it's important to
note that that this is harder in written text hypothetically
(13:09):
than it is in spoken text because of all of
those clues that we get from each other when we're
talking out loud. Sure, I mean, it's much easier to
detect sarcasm in person. I mean, you've probably had this experience,
even though you're not a computer. You are a real
human I assume, Yeah, we're guessing at least half of
you are. Assuming you are a real human, you have
(13:30):
probably had instances before where you can't read sarcasm in
an email Why why is this person so angry at me?
And then you need to include a winky face right there,
like I was just joking, and then you feel foolish.
But I mean, the lack of all this context that
we get from body language and tone and stuff can
be difficult, and in fact, that's part of the application
(13:54):
of this, but I'll get into that in a second.
So essentially, what this tone and lizer is doing is
it's it's examining all the words within that block of
text and tagging those words based upon how a predetermined
categorization has been set up for the tone analyzer. So
certain words fall under the general category of cheerful. That's
(14:18):
a great example because it pops up a lot and uh.
Then it ends up giving you a percentage of the
overall tone of the piece. You can actually do it
in two ways. You can get a word count or
you can get a percentage approach. The word count just
tells you how many of the words within a given
block of text fall into each category, So whether it's social,
(14:40):
or it's a writing que or it's an emotional cue.
And then the percentage gives you kind of a what
what is the thrust? Like? What was the impact on
the reader? So that suggests that there are waitings to
these words, right, that certain words are weighted as being
more powerful. I found the percentage results to be more
interesting because the word count results just tell you most
(15:03):
of your words or social orientation, which could be simple
things like the and but or like that doesn't have
that doesn't impact a lot unless you're using them really well.
But the words that were or butts can be extremely
extremely uh powerful. Yes, yes, but but Ideally, what this
(15:29):
would allow you to do is tell how your work
might impact a reader before the reader has a chance
to to see this written work, and you might be
able to then revise that work if it doesn't seem
like it's getting across the message you had intended. So
(15:49):
it's it's kind of like a hey, listen, uh, I
know you think you might be coming across this way,
but this is analytically how you are coming across. Do
you want to option to change up some of this stuff? Yeah,
maybe you could change some of these aggressive words to
be more friendly, for example, yeah yeah. Or it may
be that if it's in a business letter, maybe the
(16:10):
aggressive tone is exactly what you wanted. So it gives
you the opportunity to go the other way. And in fact,
what will happen is when you use the the app,
it highlights each word or most of the words, almost
all of them in a block of text, and then
by clicking on it, you can have options to change
that word to to a synonym, and it even categorizes
(16:32):
the synonym according to the various social writing or emotional cues.
So if you wanted to make it a more emotional plea,
you could do that by going under the emotional category
and choosing a synonym under there, and presumably this would
make your presentation or letter or whatever it is more
(16:52):
have more of an emotional impact. I didn't I didn't
understand some of the options it was giving me with
the synonyms, because at one point it's suggested to make
my thing more agreeable, I changed the word bad to
the word lousy. Yeah. Yeah, I mean, well it's more
jokey like, kind of like like it's got I think
(17:12):
that lousy has a slightly softer connotation. Lousy just bad,
I would say, comes across as a little more playful, right,
Like if I were to say Lauren is a bad
person versus Lauren is a lousy person. Yeah. Yeah, you know,
if you're just a lousy person, then you know you
can pick those louses off of that shampoo. Yeah, much
(17:36):
better than being bad. Yeah. It reminds me of that
amazing Michael Jackson song Lousy video for that was amazing.
So the idea of being with a tool like this
in the future, you could actually build this functionality directly
into something else like a word processor or an email
program or ta most useful. Right, the idea being that,
(18:00):
or you save a document before you send an email,
before you send off that text message, or update your
Facebook post. The analyzer could tell you, oh, hey, by
the way, you're coming across as a told jerk face,
this should be beside the tweet button. Yes, it should
be well and I think, you know, making it something
that is um like word count or spell check, spell
(18:22):
check or even grammar check these days. Yeah, I think
as long as it didn't get in the way of
what you were doing, if it were an optional thing
for you to check, or it was you know, unobtrusive
and be great, underlined it in in fuchia or something
like that, you know, whatever color is left over, right,
so exactly something that's not going to easily be confused
with the colors were already using for everything else with
(18:42):
grammar and spelling errors. But it would be helpful, and uh,
it might mean that you would avoid some of those
situations where you dash off a message to someone thinking
it's perfectly fine and they receive it and are offended
or confused or or hurt or whatever that may be. Well,
let's talk about some examples. Yeah, so there is a
demo available online. Uh, you can go to the tone
(19:04):
analyzer website and actually try the live demo. It allows
you to paste in or type in a block of
text and then you can analyze it um And so
I decided I would throw in. It took me a
long time to decide which opening line I would use
from all the different novels that I love, but I
chose this one. See if you recognize it. It is
(19:25):
a truth universally acknowledged that a single man in possession
of a good fortune must be in want of a wife. Yes,
is from Black Books. The series actually is used in
Black Books, but Black but that's not obviously where it's from.
At any rate, it's Jane Austen quote obviously. So the
(19:45):
analyzer said that the emotional tone was cheerful, the social
tone was open and agreeable, and the writing tone was analytical. Uh.
Looking at a word count like Joe was saying, not
as interesting. So I just went with the percentiles. But
if you did look at the word count, only five
percent the sentence is in an emotional tone, in social
(20:09):
tone and five percent in writing tone just by words.
But when it looks at the impact, the emotional impact
and the UH and the social um or the writing
impact rather is much greater than the social impact. So
in that case it was almost like a third or
third and a third. So even though there were fewer
writing cues and emotional cues in the sentence I picked,
(20:32):
they had a greater impact than the social cues did,
which I thought was kind of interesting. But it also
shows that when you when you do this, you see
how how the analyzer is picking out each word and
classifying it. Yeah, it breaks down by color. Yeah. Yeah,
So if you are looking at the emotional words, those
are highlighted in shades of red or pink. Uh. And
(20:53):
you know, the the emotional one is divided up a
little bit too, so that way, like cheerful is one
of them, and I think that's a very pink color
that they used for cheerful, And so all the cheerful
words were bright pink um. Then social words were all
in shades of blue, and writing tone words well in
shades of green. And if you clicked on them, that's
where it would give you the option to switch those out,
(21:15):
so that you can you know, if you did you
did you really mean in want of a wife? Uh?
But also I should point out that the results we
get are you know, one thing you have to keep
in mind is what is the basis of comparison, because
it's not just a universal you know, text analyzer, it's
(21:35):
actually analyzing it against a standard. And in this case,
the standard they were using was the standard you would
use for business letters. So according to you know, compared
to your business letters standard, Jane Austen, you know, it's cheerful,
so I'll buy that. But yeah, it's one of the
other things is that I think a full tone analyzer
(21:56):
would have different comparisons you could use, not just the
business letter roach, it would also like compare you to
threatening letters from creditors. Yeah, and and maybe maybe we
can finally get that direct computer comparison of Jane Austen
to like Hemingway or something like that. It would be interesting. So,
you know, I know there was an article in which
(22:16):
an author, uh or the writer of the articles said
that they had compared themselves to Mark Twain. Oh yeah,
I read that, yeah, which was entertaining, and said that
according to the analysis, the two wrote in a very
similar way, like the percentages came out in a similar way,
and actually used that as a means of talking about
the limitations of that but Joe, you used an example
(22:39):
that also kind of showed some limitations of the tone analyzer. Yeah.
I was like, well, what's it gonna make of something
really philosophical? So I put in a quote from Dustevsky's
Notes from the Underground. Okay, cheerful stuff that the underground
man is talking in his long sort of diary section
in the first half, and he says I could not
(23:02):
become anything, neither good nor bad, neither a scoundrel nor
an honest man, neither a hero nor an insect. And
now I am eking out my days in my corner,
taunting myself with the bitter and entirely useless consolation that
an intelligent man cannot seriously become anything, that only a
fool can become something. All right, So what what did
(23:25):
the analyzer have to say about this? Well, again, I'm
going with the percentile feedback, not the word counter. This
is more interesting. It says the emotional tone was anger
one percent seems accurate. Negativity one okay, yeah, I got that,
and cheerfulness cheerfully angry and negative. Actually that's not a
(23:47):
bad interpretation that passage. It is kind of manic in
a way. Yeah, yeah, it's it's gonna upbeat about the
fact that that the world is bleaking dire and there's
absolutely nothing to be done about to make me wonder
what the five emotions in the head of the underground
man all inside out would be doing at this particular point, right,
(24:08):
So I want to say more about that in a minute,
but just a little more on the result. Probably I'm
guessing you're right. Yeah, but yet he refuses to do
anything about his liver problem out of spite anyway. So
it also says among the social tone breakdown, it was
fort agreeable, nous, zero percent conscientiousness, and zero percent openness.
(24:32):
And then the writing tone was a hundred percent analytical,
zero percent confident, and percent tentative. And I was like, well,
that's about zero percent confident. Is this is really telling
a description of notes from underground? So well, no, it's
actually like simultaneously a hundred percent confident and zero percent confident. Uh,
(24:53):
tentative might be a good word. So anyway, I highlighted
four words, and that feedback the it are, so it
tells you why it's rating in a certain way. So
I was looking for the word did that cheerfulness result
come from? And it was like, look, you said the
words good, honest, hero, and intelligent. Clearly you were being cheerful, right.
(25:17):
I mean, this is hilarious if you consider the passage
in context, because the first three of those good, honest,
and hero are counterfactual negations. He says, I am neither
good nor bad, and I'm not these things. And then intelligent,
he's not saying like it's good to be intelligent. He's
talking about it being a curse to be intelligent. Yeah,
(25:40):
and this actually highlights a problem with the analyzer in general,
which is that it's looking at individual words, but it
cannot necessarily understand the context of those words. Yeah. Yeah,
it's doing nothing to part the context, right. It's it's
it's just isolating each word and then kind of doing
a tally at the end and saying, well, here are
(26:01):
all the good words, and here are all the negative words,
and here are all the you know, the the indifferent words.
And when you weigh them all out by there the
there how many there are and the actual emotional weight.
I don't know what the waiting system is, but it's
clearly not a one to one sort of thing. But
once it all figures it out that here's the result,
So it would probably give the cheerfulness thumbs up to
(26:24):
an email that I sent to my boss if it
said like, you are not good, you are not honest,
and you should not go on living. Yes, exactly, if
you were to type I am glad and then you
typed I am not glad, you would get very similar
results in the tone analyzer because it's picking out glad
as being a cheerful word, but it's not figuring out
(26:46):
that one of those senses essentially is the opposite of
the other. You know, it doesn't know that. Uh So
perhaps maybe one day the tone analyzer will actually be
able to understand context as well, beyond on just these
these you know, recognizing the individual words, but understand what
collectively they are trying to get across, so that when
(27:07):
you do analyze the tone of a message, it's more accurate.
So before you send that email, before you save that document,
before you you know, deliver your presentation, you can get
a cold unfeeling robot to tell you how much of
a cold unfeeling robot you are. Cold unfeeling robot could say,
you know, based upon the analytical uh or based on
(27:30):
the data analysis of how what your word choice and
the way you put them together, you're going to come
across as a real minch. I mean, you know, are
you sure you want to keep comparing your coworkers to insects? Mr? Kafka?
Are you are you sold on this being a bug thing? Uh? Yeah?
(27:51):
And until until that point, it's really just kind of
a interesting footnote in in the overall history of I
B M and they're wonderful word processing technologies. They actually
coined the term word processing back in nineteen sixty four
in the marketing materials for an electric typewriter, the first
(28:12):
electric typewriter that had a magnetic tape memory drive. Ye,
my dad had one of those. Uh not that, not
that particular generation of word processors, but dad did have
one was of that same Before we had a computer,
we had a word processor. Um. And it also illustrates
how difficult natural language and understanding humans. Uh this how
(28:34):
difficult that problem is for for computers. For artificial intelligence. Uh.
We we see a lot of developments in AI that
are really really promising, but we have to remind ourselves
there's still a long way to go, and there are
other things that we can talk about. Two, this isn't
the only tool that's ever been built to try and
understand the tone or whether or not someone's being sincere
(28:56):
in a message. Right. Oh of course. So back in
there was a group of researchers out of the Hebrew
University in Israel who designed Sassy. That's the semi supervised
algorithm for sarcasm identification, and that's so awesome. It's even
better because it's not Sassy with the y, it's s
(29:16):
a s I yeah. Actually, really there should be a
heart over the eye. I think so, I think so.
I think that's probably if Sassy had hands, that's how
it would write its own name, right, but it would
be a sarcastic herd over the eye, an ironic one.
Hell is sarcastic. So so this so this team set
Sassy to I'm gonna giggle a recent goole time to
(29:37):
say that name. Okay. They set Sassy to analyze collections
of five point nine million tweets and sixty six thousand
product reviews from Amazon and Okay, Since, as we have discussed,
sarcasm is most naturally conveyed via vocal tone and nonverbal cues,
they first had to map out what sarcasm actually looks
(29:57):
like in text and and came up with kind of
kind of a matrix of of like hyperbolic words and
excessive punctuation or using lots of ellipses in particular, was
something that they earmarked and straightforward sentence structure diagramming sarcasm. Yeah,
and I think this was probably the most difficult part
(30:18):
part of their research, like everything else after that, Like
like creating these patterns for the machine to look for
was the most difficult thing. So they gave it examples
of sarcasm by by feeding it tweets that were tagged
like hashtag sarcasm and also one star Amazon reviews that
had been deemed sarcastic by a panel of fifteen humans.
(30:39):
It's funny because they could have just had the computer
follow certain Twitter accounts and say, like, you can be
any tweet coming from this account is sarcastic. Also, remember
this was back in so so Twitter was relatively new
at the time. Yeah, knew enough that in their report
about it, they spent like a long time describing what
(31:01):
Twitter was. Uh Okay, So they then instructed Sassy to
rate sentences from one to five, with one being not
sarcastic at all and five being super sarcastic, and they
found that Sassy could identify sarcastic Amazon reviews with precision,
and it did even better on Twitter, kind of unexpectedly
(31:24):
because there's less context to work with in in tweets,
which are short and kind of stream of consciousness. A
lot of the time it's precision rate over there was seventy.
When you're limited to characters, it is is something of
an art to get across sarcasm in a way that
people realize it's sarcastic. By the way, not definitely an art,
(31:45):
not a science. Oh yeah. I have posted many things
where I thought, well, clearly people will understand that this
is not an actual sincere statement, and I have been wrong.
At least one person proves me wrong. Well, this will
be the second time in a couple of weeks that
I've had to bring up pose law on a podcast.
I mean, if you're saying something that sounds extreme to
(32:08):
parody extremism, people will take you seriously because it's hard
to tell. It's hard to tell the difference between a
parody of an extreme view and an actual extreme view, right, yeah, yeah,
That's why a lot of the kind of rival sites
that are popping up to the to the onion. I
think fail really hard sometimes because I'm like, oh, you
guys need to take it much further or just stop
(32:30):
writing entirely, because satire just comes across as a lie, right, yeah,
and slander is different things than satire, Yes, so, but
so so. This research team decided that Sassy was really
good at detecting very straightforward sarcasm, the kind that in retrospect,
lots of people probably use on Twitter, precisely because you
(32:53):
have such limited space and limited context, so you have
to be pretty direct. Sassy made a lot more falseness
of evaluations than false positives, which indicates that a lot
of the more subtle stuff or the more intricate stuff
was slipping by it. Okay, so could you give an
example of, like the difference between straightforward sarcasm and non
(33:15):
straightforward sarcasm? Uh? Yeah, sure. A simple one might be
something like wow, Mondays are my favorite days ever? Exclamation point,
exclamation point, exclamation point, and something a little bit more complicated.
I can't say whether or not this was actually used
in the study, but I cribbed this from um amazon
(33:36):
dot com from from that banana slicer. If you guys
have seen that. Yeah, one of the reviews is I
tried the banana slicer and found it unacceptable. As shown
in the picture, the slicer is curved from left to right.
All of my bananas are bent the other way, so,
you know, a little bit harder for a computer to
pick up on that, right. Yeah. There are other accused
(33:56):
as well that they would rely upon, like the hashtag sarcasm.
If there's certain phrases that without that hashtag, they're not sarcastic,
like I can't wait to get home tonight. If I
just post that, then it seems like I can't wait
to get home tonight. But if I do hashtag sarcasm,
then you're vague tweeting and it's annoying. Well, no, then
you know I'm you know I you know I definitely
(34:18):
can wait before I get over tonight. You might not
know why, but you know that I'm not looking forward
to it, right, which is vague tweeting, which is annoying.
It's not as vague as some of the stuff I
see on there. At least with the hashtag you realize
what the tone is. It's not as vague at any rate. Yeah,
So another example of this going back to IBM and
(34:39):
and Watson. Actually, UM is the use of the the
that language recognition, that natural language cognitive computing approach toward
perhaps like getting computers to uh you come up with
some arguments. Yeah, this is an interesting thing that I
read about last year and wrote a blog post for
(34:59):
a website about is the Watson debater. Yeah. So this
was a new iteration in the development of Watson technologies
that was designed to look for arguments like statements in
support or in opposition to a proposition. And this is
(35:20):
interesting because it's more difficult than you might think to
do this. So there was a presentation about it last
year at the Milken Institute Global Conference. The Milken Institute
is an economic think tank. I didn't know that. I
looked it up. I was curious, but uh yeah. The
conference had a whole bunch of different people give a
(35:41):
presentation about the UM the near future and it was
kind of like you know what's next gonna is gonna
blow your mind, kind of sort of the same sort
of stuff we like to talk about here on Forward Thinking.
And they had several guests. Among them was a representative
from IBM, said John Kelly. Yeah, and and he was
unveiling the debate of the first time to the public.
It was something that up to that point had only
(36:03):
been discussed internally at IBM. Right, and so what he
was demonstrating was the debater's ability to not come up
with arguments because we're nowhere near there yet that level
of processing and synthesis. But it could look at a
whole bunch of articles and say, what are some statements
(36:24):
in support of or in opposition to a proposition. So
they gave the example and the presentation of the proposition
the sale of violent video games to minors should be banned.
I love it because it includes banned Yeah, and and
and it's it's cool. The process that this this tool
uses is pretty interesting. First, it scans pretty much everything
(36:47):
and has access to UH and in this case the
demonstration they showed at the conference, it scanned four million
articles and this was just to see if the articles
that it had access to were relevant to the actual question.
Then it picks some top contenders, yeah, the ten best
fits to the topic at hand, and then scanned all
(37:11):
the sentences in those ten best articles a yeah, going
through three thousand sentences, and then started to classify those
sentences as being either in favor of banning violent video
games or against the banning of violent video games at
least the sale of violent video games to minors. And UH,
(37:33):
they looked for sentences that contained what they called candidate
claims that would either be one of those statements for
or against the central premise. UH, and then they they
it would identify the parameters of those claims and then
assessed those claims whether or not they were truly pro
or con and then put them into those those camps.
(37:55):
So ultimately you would ask the computer, all right, so
what are the arguments for and against this proposition? And
it would say, in favor of the motion, violent video
games should be banned, it should be noted that violent
video games actually cause aggressive behavior. And then you know,
some other statements along those lines, and then it would say,
(38:16):
in opposition to the proposition violent video games should be banned,
it should be noted that violent video games do not
actually cause violent behavior. Right, there's no causation link. Yeah,
which is funny because it'll it'll take to completely contradictory
statements and cite them both. Right. Yeah, there are certain
things that that leap out at you, especially with this
(38:37):
particular example. And it's funny because I watched it and
I made notes, and then I read the rest of
your blog and posting, and you and I have the
exact same reactions. Yeah, one of them being that one
of the statements that just rubbed me the wrong way
was about how it was essentially said that video games,
the playing of video games is a a popular pastime
(38:58):
for boys. It says playing violent video games is part
of a boy's natural childhood development. And it was just like, Okay,
I know the computer is not sexist, but it just
has access to some sexist views. And this this shows
the limitations of this approach, right, oh sure, well and
and and it shows really the limitations of humans at
(39:19):
that point to to a to okay, yeah, well, to
to come up with convincing debate or convincing debate statements
in in saying something completely unsupported and also in just
making making stuff up. Yeah, so yeah, like why, I
don't know what the funny thing to me was. It
(39:41):
had plenty of statements it could have chosen from. I like,
actually found the Wikipedia article that it was cribbing its
arguments from, and I looked at the article and I
was like, there were better statements in this exact same
paragraph you could have used than that one. Why did
that statement get elected? And ultimately, I don't know. One
(40:02):
thing we we don't necessarily know is when it looked
at the Wikipedia article, which could have been altered between
the time it did its research and when you did yours.
But I doubt that it was that significant if it
was within that same paragraph, I doubt there was a
lot of there were a lot of changes. I mean,
if that particular sentence didn't appear at all within the paragraph,
then I would say, oh, something's happened. But at any rate, yeah,
(40:26):
it's it's it was problematic with that particular example, but
it does show that it's reliant completely upon the It's
relying completely upon the content that humans have created undeniably
since the beginning of time. It has been a natural
part of boys child, since we dwelled in the halls
(40:50):
of Castle wolf in Stein. I'll never forget the uh,
the story I read about the ancient humans playing Wooly
Mammoth Destroyer and uh, you know games like that. So
but no, I mean, yeah, you're exactly right. It's just
reflecting the fact that all it has to work with
(41:13):
is what humans have said. And it doesn't have a
good criteria as far as we can tell for judging
the difference between a good supporting argument and a bad
supporting argument. Interesting, it can't parsing for is this in
support or not? Yeah, it can't. It can't judge is
this a good logical argument? Right? It could actually come
(41:34):
up with arguments that contain logical fallacies in them. That's
the really interesting thing is that. And it's not because
the that's not doing what's supposed to do. It's doing
exactly what it was meant to do. It's just that
the material it's pulling from itself is flawed. So, yeah,
we can't think of it as some sort of magical
oracle that we can consult. And it has access to
truth with a capital T. It has access to the
(41:56):
same stuff you and I would have access to if
we just poured over Wikipedia for a few days when
we were trying to research something and had no judgment criteria. Sure,
but but it's you know, hopefully in the future or
possibly in the future if they choose to continue developing it,
it could turn into, if not an oracle, at the
very least a source of of better advice well, and
(42:17):
and the thing that they were talking about using it
for didn't have anything to do with trying to figure
out a debate like a political debate or or a
social debate. They were actually talking about UH using it
in the context of UM again, the discipline of medicine,
where let's say that you are part of UH a
(42:38):
hospital administration staff, and you're trying to determine which policies
and procedures you want to put in place in your hospital,
and so you have to do all that research. You
have to figure out what are the benefits, what are
the drawbacks of all these different alternatives, and potentially you
could use UH technology like this that would then come
back with the pros and cons of each approach, allowing
(43:00):
you to make a more informed decision. I'm sorry, I'm
just loving the idea that the machine that identified playing
violent video games as part of a boy's natural childhood
development deciding what anesthetic does I get To be fair,
it's not relying on Wikipedia in the second instance. Okay,
(43:21):
all right, so we can relax on that now when
we get the home version there and ask which demonstrable,
which transform is demonstrably the most important, and then it
would be able to tell us. Yeah, I mean obviously,
I don't think doctors are actually just going to seed
all of their authority to the Watson debater and say, well,
what do we give them? You know, they were they
were specifically using this as the demo of what the
(43:43):
technology can do, not as a you know, this is
exactly what's going to rely upon. So instead of relying
upon Wikipedia, to be relying upon the literature in the
medical field, the research that's been done. And even then
I assume it would be in an advisory function nothing decision.
It would essentially well, again, if it's getting pros and cons,
then ultimately the decision still falls on the humans. It's
(44:04):
not able to wait which ones are more important. Like
it could say the pros are your hallways will be
less cluttered, the cons are seventy of your patients will die,
and like, I recommend you do this because you have
cluttered halls. That's not gonna be how it turns out. Well,
I can see something like this actually being useful to
me in my job, just as a shortcut for digging
(44:24):
up leads on a subject. For example, like you know,
I read an article about a subject I've never read
about before, and the article has one pretty clear view
on it, like it's in favor of X, and I
don't even know what the arguments against X are. If
there are any something like this could be really useful, like, oh,
here are some things to look up. I think it
(44:46):
would be really useful for something like, let's say that
we wanted to cover a controversial topic in physics, something
like because I've seen it pop up yet again. The
m drive would be a great example. Something that could
scour through the actual scholarly material and come back with
the pros and cons of something like that, so that
(45:07):
we at least have a starting point. And I think
you can go look up the what's behind the exact
kind of like the references in Wikipedia. Right, you go
to a Wikipedia article, you scan down to the references,
and you go to those sources to see if in
fact they do reflect what is or if the article
reflects what was in the actual primary sources, which is
the proper way to use Wikipedia. Yeah, yeah, I agree. Yeah,
(45:29):
And it kind of funny how often that's a dead
link and you don't know if it's ever been alive. Yeah, Well,
you can always go to archive or dot Oregon and
check it out. I've done that many times. Actually. Uh So,
one of the things I wanted to talk about really
briefly to kind of wrap this up, is is this,
actually the future of computing is cognitive computing. This idea
of creating a machine that can, at least on some level,
(45:52):
mimic the way we humans think is that the future.
And it seems like more and more people are leaning
towards Yes, that is the future. But it is the future.
It's not now. Yeah, it's it's one of those I mean,
I don't know if you guys can tolerate the theme
song today, but but it's it's going to happen in
a certain period of years from now, right, so that
(46:18):
the neural network software, the hardware that's necessary it it
requires a lot of processing power in order for it
to even come close to what we humans are capable
of doing just naturally. And in fact, in that same
presentation that that Kelly did at the conference, he pointed
out that brains are really amazing and said that you know,
(46:41):
they require very little power. It's the equivalent of about
twenty watts, so less than a lightbulb. Most lightbulbs UM
and if you compare that to a supercomputer, it's nothing.
And you also look at the intricate connections that our
brains have. We have billion neurons and then all the
different connection is between those neurons, trillions of them. It's
(47:03):
it's phenomenal what our brains are able to do. Meanwhile,
if you wanted to create a simulation of that, they
talked about how they used UM a supercomputer at the
Livermore Laboratory that was a simulating a neural network like
a human brain, and it required eight megawatts of power,
So eight million, what's as opposed to twenty with one
(47:27):
point five million processors, so compared to eighty two million neurons,
UH capable of processing six point three billion threads simultaneously.
And it ran hundred times slower than the human brain.
So just simulating a fraction of what our brains can
do requires much more processing power, much more literal power
(47:51):
like electricity, and it takes longer. So while this is awesome,
and we love the idea of having computers that can
actually learn as a post to you have to program it.
You know, if you can teach the computer something over time,
and we this goes back to the Google Dream stuff
we were talking about two. Um, that's fantastic, but we're
still at a point where until there's a breakthrough, either
(48:14):
on the logical side or the programming side, they're lagging
way behind what we we humans can do, just because
we're fundamentally different. But it is exciting that this is
a trend, and I expect we'll see that trend continue.
And I love the idea that at some point I'm
going to have a little app on my phone that
will warn me before I send a text to my
(48:35):
wife so that I can reword it and she won't
be mad at me for for good reason. But you know,
it might be a misinterpretation but accidental. But it falls
on my shoulders that I did not word things in
a way that misinterpretation. Did you? Did you really want
to say that you're a lousy wife? Yeah, maybe you
(48:56):
want to say bad. No, I don't want to say that, thanks,
But yeah, this was this was fun to look into. Uh.
I recommend you guys, if you haven't played with the
tone analyzer going go and play with that. Either take
a paragraph from someplace or type something up. Or I
even when I did the video episode about this. I
(49:18):
took the first paragraph I wrote for the episode and
fed it back through the analyzer, and it said I
was very open and conscientious and cheerful, So I got
it right. Um, but yeah, you should go check that
out because that demo is available for everybody, gold Star.
You know what's going to happen when you run your
secret underground diaries through it. It's probably going to say, like,
(49:42):
don't ever show this to anyone else. So you, guys,
if you have any suggestions for future episodes of forward Thinking,
you should let us know to analyzer. Yeah, yeah, that'd
be great. Write up an email to us, put it
through the tone analyzer, change of the words are bitrarily
and then send it along to us and we'll see
what do you think of it? Make it more cheerful? U. No,
(50:05):
but seriously, if you have a suggestion, you should write
us our email addresses f W Thinking at how Stuff
Works dot com, or drop us a line on Twitter,
Google Plus or Facebook. At Twitter and Google Plus, we
are f W Thinking. Just search fw Thinking in that
handy dandy search bar on Facebook. We will pop right up.
You can leave us a message and we'll talk to
you again. Really, soon. For more on this topic in
(50:31):
the future of technology, I'll visit forward thinking dot Com,
brought to you by Toyota. Let's Go Places