I’m a computational linguist my job is
to help computers understand language so
when I was first asked to give this talk
I wished my field was a little further
along so maybe computer could give my
talk instead of me and I wouldn’t have
to stand here and be nervous
as you can tell from the fact that I do
stand here and I’m nervous
our field is not that far along
computers cannot understand language and
that has to do with the fact that even
if they did they couldn’t replace me
here because often times it’s more
important who says something then what
is said now this might seem a little
counterintuitive after all we all have
devices in our pockets or in our homes
that we can talk to we can say hey
what’s the weather gonna be like today
or what’s on my agenda this afternoon
and they respond they talk back so how
do they not understand us plus
everything we think is hard about
learning a language
remembering vocabulary cramming in these
weird declension tables those are all
things that are dead easy for computers
however language is a lot more than just
learning vocabulary and declension
tables language is fundamentally a human
endeavor and no matter how much these
machines learn they cannot make up for
that so if I say something like I went
to Pittsburgh last month and then I
asked did my head go to Pittsburgh as
well you would think that’s a dumb
question because without head it’s very
hard to go anywhere but for any of these
tools that’s a really difficult question
to answer they have never seen anything
like that in their training data and so
it’s really hard for them to tell what’s
going on the thing is that language is
fundamentally about ourselves
we express who we are through language
in subtle and not so subtle ways and we
account for that when we encounter
somebody and talk to them so within a
few sentences you can typically tell
roughly how old the person is where
they’re from probably their age their
gender their personality their attitude
a lot of things but to a computer these
things are not obvious they don’t take
them into account they only pay
attention to what the people say but not
who says it and that does have an effect
on understanding so if you see a
sentence like that was a sick
performance it makes a huge difference
whether that was said by a 16 year old
at a skate park or by an 86 year old
after a piano recital and we take that
meaning difference into account because
when we hear something we also listen to
who says it but computers do not and
that creates a problem now this comes
from mainly how we approach language
right now in computers namely as an
engineering problem and that has had a
lot of success right now you can log on
to the internet and you can translate
from Italian to English to Chinese to
German to whatever language you choose
and it’s actually not half bad you can
run a computer and find out which tweets
are good or bad for your company you can
have a computer analyze better than any
human could what the verb is in a
sentence but this all relies on
recognizing patterns we feed computers
with a lot of examples and ask them to
find regularities in there and then
respond in a predefined way what we’ve
done with that is to basically take
something that’s beautifully complex
like language and made it into a
relatively simple straightforward
formation theoretic measure and like I
said that helps and a lot of smart
people have worked on making it so but
it doesn’t really help us make computers
understand language the problem is that
what we teach the computers stays with
them for their entire life so most of
the data we give them most of the
examples we present to them are actually
pretty dated a lot of current language
technology relies on newspaper articles
that were written in the 1980s very few
people here probably speak like
journalists from the 1980s and that’s a
good thing but that means language
evolves language changes but not for the
computers who have learned on the 1980s
data this is a bit as if you grew up in
a place where you only heard dialect
spoken and then at some point you go out
into the world and you meet people who
don’t speak dialect you would have a
hard time understanding them and that’s
exactly what happens with computers they
try to match everything to newspapers
from the 1980s and more often than not
they fail we’ve actually seen that
computers are incredibly bad at
analyzing language from people who grew
up after the 80s and they’re incredibly
bad at analyzing language if it’s
written in a slight dialect or with
variations now what’s the problem with
that so what computers aren’t perfect
well there are several problems
computers are tools we want to use them
to actually do things and they’re
ubiquitous tools chances are if you’ve
logged onto the internet today or you’ve
pulled out your phone you’ve interacted
with language technology and if that
language technology doesn’t really work
because the computer doesn’t really
understand you then it’s a broken tool
now if you’ve ever seen a
left handed friend try to use a pair of
scissors made for right-handed people
you’ll know that a tool that wasn’t
designed with you in mind is actually
extremely awkward inefficient and
potentially dangerous
why dangerous well it turns out there’s
a lot of left-handed speakers out in the
world and they all can’t benefit from
the language tools that we have today
and these tools are increasingly used in
politics in industry to analyze new
hires to analyze business strategy to
analyze things in almost every realm of
life and if these tools aren’t perfect
we risk actually leaving out a lot of
left handed speakers the thing is that
it’s actually easy to address we’ve
shown that if you teach a computer that
different people use language
differently they can perform a lot
better they can learn that the
difference between how young and old
people and men and women speak is
actually beneficial in identifying what
a text is about different people talk
about different things in different ways
they can even learn better whether
something is positive or negative this
becomes especially interesting when we
look at the intersection of language and
personality we know that people who
suffer from mental health conditions
express them through the way they speak
consciously and subconsciously and
psychologists will pick up on that in an
interview with a patient
however psychologists cannot be with
their patients all the time they often
only have 15 minutes or something like
that in order to make a potentially
life-changing decision for somebody if
we could develop tools that could stay
with those patients and listen to them
and pick up on signals that show that a
depression or suicidal
is imminent they could actually help us
prevent these things and we know that
these tools exist computers are really
good at picking up on subtle things and
making a decision based on these
patterns however all of these tools that
have been developed suffer from the
whatnot who trap they upend pay only
attention to what is being said but not
who says it and especially in this case
that makes a huge difference we were
actually able to show that if we teach a
computer that men and women with for
example depression or suicidal thoughts
speak differently about it these tools
perform significantly better than tools
that haven’t been taught about the
differences and that makes a difference
in real life
in the experiments we ran the difference
was between one hundred and twenty
people correctly identified for a risk
of suicide or not only by paying
attention to who says something and that
is a positive message now
we’re not all the way there yet we still
have a long way to go in order to really
make computers understand us but if we
actually manage to teach computers that
who says something is just as important
as what is said we can actually make
computers into the tools that we want
powerful tools that help us in all walks
of life and that can be used equally by
everybody and not just right-handed
speakers thank you you
to help computers understand language so
when I was first asked to give this talk
I wished my field was a little further
along so maybe computer could give my
talk instead of me and I wouldn’t have
to stand here and be nervous
as you can tell from the fact that I do
stand here and I’m nervous
our field is not that far along
computers cannot understand language and
that has to do with the fact that even
if they did they couldn’t replace me
here because often times it’s more
important who says something then what
is said now this might seem a little
counterintuitive after all we all have
devices in our pockets or in our homes
that we can talk to we can say hey
what’s the weather gonna be like today
or what’s on my agenda this afternoon
and they respond they talk back so how
do they not understand us plus
everything we think is hard about
learning a language
remembering vocabulary cramming in these
weird declension tables those are all
things that are dead easy for computers
however language is a lot more than just
learning vocabulary and declension
tables language is fundamentally a human
endeavor and no matter how much these
machines learn they cannot make up for
that so if I say something like I went
to Pittsburgh last month and then I
asked did my head go to Pittsburgh as
well you would think that’s a dumb
question because without head it’s very
hard to go anywhere but for any of these
tools that’s a really difficult question
to answer they have never seen anything
like that in their training data and so
it’s really hard for them to tell what’s
going on the thing is that language is
fundamentally about ourselves
we express who we are through language
in subtle and not so subtle ways and we
account for that when we encounter
somebody and talk to them so within a
few sentences you can typically tell
roughly how old the person is where
they’re from probably their age their
gender their personality their attitude
a lot of things but to a computer these
things are not obvious they don’t take
them into account they only pay
attention to what the people say but not
who says it and that does have an effect
on understanding so if you see a
sentence like that was a sick
performance it makes a huge difference
whether that was said by a 16 year old
at a skate park or by an 86 year old
after a piano recital and we take that
meaning difference into account because
when we hear something we also listen to
who says it but computers do not and
that creates a problem now this comes
from mainly how we approach language
right now in computers namely as an
engineering problem and that has had a
lot of success right now you can log on
to the internet and you can translate
from Italian to English to Chinese to
German to whatever language you choose
and it’s actually not half bad you can
run a computer and find out which tweets
are good or bad for your company you can
have a computer analyze better than any
human could what the verb is in a
sentence but this all relies on
recognizing patterns we feed computers
with a lot of examples and ask them to
find regularities in there and then
respond in a predefined way what we’ve
done with that is to basically take
something that’s beautifully complex
like language and made it into a
relatively simple straightforward
formation theoretic measure and like I
said that helps and a lot of smart
people have worked on making it so but
it doesn’t really help us make computers
understand language the problem is that
what we teach the computers stays with
them for their entire life so most of
the data we give them most of the
examples we present to them are actually
pretty dated a lot of current language
technology relies on newspaper articles
that were written in the 1980s very few
people here probably speak like
journalists from the 1980s and that’s a
good thing but that means language
evolves language changes but not for the
computers who have learned on the 1980s
data this is a bit as if you grew up in
a place where you only heard dialect
spoken and then at some point you go out
into the world and you meet people who
don’t speak dialect you would have a
hard time understanding them and that’s
exactly what happens with computers they
try to match everything to newspapers
from the 1980s and more often than not
they fail we’ve actually seen that
computers are incredibly bad at
analyzing language from people who grew
up after the 80s and they’re incredibly
bad at analyzing language if it’s
written in a slight dialect or with
variations now what’s the problem with
that so what computers aren’t perfect
well there are several problems
computers are tools we want to use them
to actually do things and they’re
ubiquitous tools chances are if you’ve
logged onto the internet today or you’ve
pulled out your phone you’ve interacted
with language technology and if that
language technology doesn’t really work
because the computer doesn’t really
understand you then it’s a broken tool
now if you’ve ever seen a
left handed friend try to use a pair of
scissors made for right-handed people
you’ll know that a tool that wasn’t
designed with you in mind is actually
extremely awkward inefficient and
potentially dangerous
why dangerous well it turns out there’s
a lot of left-handed speakers out in the
world and they all can’t benefit from
the language tools that we have today
and these tools are increasingly used in
politics in industry to analyze new
hires to analyze business strategy to
analyze things in almost every realm of
life and if these tools aren’t perfect
we risk actually leaving out a lot of
left handed speakers the thing is that
it’s actually easy to address we’ve
shown that if you teach a computer that
different people use language
differently they can perform a lot
better they can learn that the
difference between how young and old
people and men and women speak is
actually beneficial in identifying what
a text is about different people talk
about different things in different ways
they can even learn better whether
something is positive or negative this
becomes especially interesting when we
look at the intersection of language and
personality we know that people who
suffer from mental health conditions
express them through the way they speak
consciously and subconsciously and
psychologists will pick up on that in an
interview with a patient
however psychologists cannot be with
their patients all the time they often
only have 15 minutes or something like
that in order to make a potentially
life-changing decision for somebody if
we could develop tools that could stay
with those patients and listen to them
and pick up on signals that show that a
depression or suicidal
is imminent they could actually help us
prevent these things and we know that
these tools exist computers are really
good at picking up on subtle things and
making a decision based on these
patterns however all of these tools that
have been developed suffer from the
whatnot who trap they upend pay only
attention to what is being said but not
who says it and especially in this case
that makes a huge difference we were
actually able to show that if we teach a
computer that men and women with for
example depression or suicidal thoughts
speak differently about it these tools
perform significantly better than tools
that haven’t been taught about the
differences and that makes a difference
in real life
in the experiments we ran the difference
was between one hundred and twenty
people correctly identified for a risk
of suicide or not only by paying
attention to who says something and that
is a positive message now
we’re not all the way there yet we still
have a long way to go in order to really
make computers understand us but if we
actually manage to teach computers that
who says something is just as important
as what is said we can actually make
computers into the tools that we want
powerful tools that help us in all walks
of life and that can be used equally by
everybody and not just right-handed
speakers thank you you
Please follow and like us: