Press "Enter" to skip to content

A Guided Tour through the World of Text Analytics | Mark Cieliebak | TEDxYouth@Zurich


I say that many people think about or
gaming we were already about the chess
game or something robotics or even cars
driving by themselves but then I think
sorry boring stupid text
I think we’ve this for ten years doing
research and we do business and I want
to share a little bit with you with my
experience and why I’m still fascinated
in this and I will try to compress 10
years to 50 minutes which is not so easy
so bear with me extend this like a good
magician it has a lot of tricks behind
and it looks like magic but it isn’t and
I want to explain a little bit about
where the tricks are and why it is cool
and what is impressive at the front but
behind it’s basically counting so if you
can count 1 2 3 4 then you can do many
things and text abilities and it’s not
really surprising how it works so let’s
go back in time 10 years 10 years ago
2008 the album just came out well 10
years old so we were able to search
websites and I was in a small company
very long and we did rocket science we
call it at that time business
intelligence or sometimes data
webmonitor and it was really profits I
believe we gave to some communist and
what we do with it oh this is great you
can’t do that what
that’s impressive and what we did was we
counted words on the internet so we just
collected all the websites that contain
the word
have seen Red Bull and then we come to
how many websites we come and draw the
top like that and that was WOW hey
before if they might but Wow
what is the use well it could can use it
in marketing before you have to do the
survey and people what do you think
about my brand and how popular is it and
now you could go to the web and ask well
you can count it on on the fly in real
time now what was it rocket science
because it’s not so easy as tom i mean
because a if we just come at first we
need to find out what other proper
websites and then in the website we need
to find out what is the content that we
are going to look at if you look at
website like this also like this well
you have the main text here which is the
small text here and there are give all
this clutter information which you are
not really interested in your images
with advertisements your links to other
web pages and if you want to see where a
follower pepsico less mentioned well you
probably want to concentrate on this
text now if you have a text aren’t the
leg thing is you need handle which
languages because you want to apply
these language tools that we have and
they are language specific but easier
take some common English words I knew
that he did and this piece with the
language and then people right at the
end whatever they want so they
recognized I can speak English in touch
cool as mixed language and there are
many people who really write that back
because they are grown up many languages
that are more the mix English to
what happened language was split up
without text like sentences or words
because we kind of deal with ten pages
of take up tasted one one time so you
spit out the sentences by the table at
the end right yes it worked perfectly
not always and then you need to split
the sentences into words and you can
think about what is the word or a token
and there are many issues as well to
find the proper boundary of words and
that once you have done that you have
these texts and we want to find out is
it about my brand now Coco that could
rank this by to visit their mother
planet and Switzerland which are not so
easy like solid yes and it’s not only
the spice but it’s also Salt Lake City
or the movie salt which are usually so
it leads me to some information to
figure out okay but at that time we kind
of came up with a solid for 2006 but by
2010 they were really famous the world
someday
500 million Facebook users and standard
and our customers became more and
tributon well we don’t want to have just
word counts we want to know more about
what people us
rats and one topic that came up very
often was sentiment analysis that’s
quite easy you wouldn’t know whether the
brand is a positive or negative context
because well it’s a different phone
company I mean if you come up only Pepsi
welcome guests somebody says that these
great that’s better than Pepsi thought
might a system like that so we needed to
get out of sentiment and yeah you can do
that and also can’t just come today
positive negative words and it’s easy
but as I said it was on Facebook and
Twitter
well Facebook is still okay a Twitter
makes it pretty hard so I brought you
three examples this is well it’s a
little bit terribly but try to figure
out the sentiment of these three
sentences to people may be available the
voluntary what the heck be waiting for
or this is one of your little jokes they
heaven they get the monkey next Tuesday
or you can take me if you still respect
your pets super-hard that this date
the now you’re laughing we were
supposed to find out what segment of
visas yes I mean yes and people expected
us to have a good settlement with the
currency so we asked people you want to
label this by hand okay because they
want to buy the ticket obviously a
second one and if you can figure out the
third one just let me know I don’t have
any to so madam well this was kind of
closed at some time and we spent some
time in research on that but in between
and I want to make a jump now a time
down to 2018 because in between we solve
the dollar problem and there
really the weights off by now so as I
said text justification well this is so
and Texas vacation means we can find out
or the technologies we can also find out
from tonight the age and gender of the
author
okay scary it arrived something like
twenty lines of text we can guess the
rate and gender and from which country
you come from even is the writing
English we could figure out which were
companies and the patients were
mentioned even if we didn’t have a
closed list of the company could just
say oh in this part of the text errors
of how many mentioned we could figure
out and classify which topics are there
so automatic topic generation very
sinister and we’ll talk about that later
on and machine translation well we
didn’t do it but Google did a good job
well relatively good job with our five
and very recently there’s a new system I
don’t know if they know it’s be better
which is excellent in translating and it
kind of slows down but then three years
ago people came up to the new Ramon
Ramon but externalities was dialogue
systems what is it the question is I
want to talk to my tutor and expected to
understand what I’m saying and get a
that’s really what they wanted so the
first day who were ready 2011 and that
was only what it was a huge machine
trained for one particular purpose and
it is winter aged or jeopardy
so that was game and compete against
humans in understanding reasonable
natural language tasks and I didn’t
bring the task here but but what they
did is usually Arabic we are doing
question answering there’s particular
question was with President if 2015 and
then ground first immunity from arugula
and behind it there is a lot of
Technology so what we have at the left
hand side is if you really want to talk
to you to translate your spoken language
into text it’s cool for English and
German is not so cool but with children
because there is not much per se from
all these different either makes it
different but once your text you need to
find out what is the person talking
about what is the intent what else it
does he or she want then next step is
you need to find out what is it what is
the next few are talking about so if we
look at this again I want to know who so
I want to have Percy and the entity will
feed us with president and 2015 and I’m
one who connected and find out what is
the output that match up these both for
entities and if I am more than one
hostage and reply than any player
something like cortex tracking because
if I spot talk about the video or movie
and then I are with director I expect
sister to know
movies direct direct this movie and on
the other hand well we try to create a
reply so we need to select which answer
and will be to fill in response because
the expect a reasonably pliant that’s
Alexis and then we need to put it again
into spoken language in a way for simple
dialogue questions like with the
database they are behind but Jung then
the next thing we want to check this is
not a purposeful check but we want to
politics or whatever
hi I’m sorry what’s the name Tom are you
great and so on people really want to do
that
maybe but already in Japan we’re kept
checked with your computer and so two
people who live alone to make the
custodial companion so what kind of
combination of systems were there ready
let’s do it anytime the first system was
already implemented in 1966 I chose the
fight for the Philosopher’s in the first
wave of artificial intelligence said
well we can talk to your computer and he
implemented like something like an
analysis and what he did was very simple
he took the sentence that the user input
right I have an issue with my father and
then it made a question from there why
father and people owe the cost he
doesn’t give me the car and then after
oh that’s bad why and it get went back
on and without understanding anything
people stood for up to two hours in
front of the system
typing in their personal details and in
the end when he wants to publish it they
say when you can on average it’s too
personal
so it was really successful at that time
thirty years nature internet was there
so people could say one of the most
famous systems for this you can type in
any expression any fancy one and the
what try to convince you that it is a
human and after that human and
distinguish it from any reply it has
already been in dialogue and it’s really
convincing say anything come to read the
reply deck thank you for three years ago
four years you don’t know they build
this hitchbot and his book was a cool
idea they put the technology of the
cleverbot behind and put it on the
street and said this part should travel
from Halifax to Vancouver and it should
really hitchhike so people who take it
set it into their car and that talk to
the monitor by writing and at the Sao
the year calculated battery I it arrived
after a couple of weeks and that they
did the same in the United States and it
was a couple of days and then last two
years ago say they brought up the
Ecuador where you can just ask Alexa
what is the height of it I have
authority cell phone connect we are okay
when all these systems are basically
still question reply systems very simple
nothing there before and they are moon
base so there is a lot of rules behind
human engineering well we are not work
money research is how to generate text
from scratch and this is a totally
different idea I want to just briefly
try to describe it
what if you take this a huge amount of
sentences texts arbitrary text like the
clapping is a teaching method and we
give the computer this text character by
character and it learns if there is F
then this sentence there is an elephant
and Ellis followed by highest followed
by P and so on after he learned the
entire sentence and then we give another
sentence in another sentence and it
learns like well sometimes after F there
is this L di t PE and then there is an
out of the sentence or here indeed so
you’ve done all these combinations of
characters and following characters okay
we do that for billions of sentences and
now now we can reverse it and say well
start with an S and generate the next
character that is most likely and the
next character that is most likely and
surprisingly this doesn’t output rubbish
but this of course after that
Shakespeare so the business
trained on Shakespeare nonsense texts
and if you read it oh if you were a fuel
type the courtesy of your law your
fighting Temeraire breath will rear the
gods and so on if you don’t understand
it don’t be annoyed this is rubbish text
it’s just a sequence of characters
without any meaning but I would be sure
if you put it on the stage people will
think it will be safe thank you and you
can do the same for pika prz the
domestic articles source code neatly
files you can generate music and then
what a mess is still rubbish thankfully
every with it we want that diverse
systems so we traded one dialogues movie
fact I to dialogue and that means
something strange happens it outputs
reasonably priced so if you say
something like hello hi how are you and
fine or what’s the color of the sky is
blue okay
honestly this is cherry picking it’s
always yellow or red but it says a color
of silicon understand in the way that
this is asking for a color but still
there is no way to determine what it
will say I can ask a question and I
think conversation make that relation
but there’s no way to know what it does
and this is what we are currently
working on a manky in the world in
recent energy that language processing
clear meaningful outfits and the ideas
we give it information for example about
restaurant and then it should get
generate something like the mill is a
cheap fast food park located in the city
center of neo Kafe Cecina now this
doesn’t okay
anymore but I think this is almost 7
which is here the cool stuff is this is
generated character by character so the
computer only leaves this and thousands
of examples of restaurant we use to
generated from scratch you can apply it
with any technology any same way you
have database and data subject which I
believe after computer what are you sure
yet
and then we are the most important
question do you plan on taking over the world
Please follow and like us: