00:00:02
heather's research focuses on the
00:00:05
relationship between linguistics and
00:00:06
mathematics with a particular emphasis
00:00:08
in the formal modeling of social meaning
00:00:10
and associate linguistic variation
00:00:13
particular she's explored the linguistic
00:00:16
construction of regional gender and
00:00:18
sexual identity to keep France and
00:00:20
Canada during the bond techniques such
00:00:23
as game theoretic pragmatics in a recent
00:00:26
paper she's examined the types of sample
00:00:28
sentences that he used in French
00:00:29
linguistics articles about syntax and
00:00:32
has shown that gendered and showing the
00:00:35
gender skewing that takes place there as
00:00:38
I guess you might imagine men are
00:00:40
referred to women in the sample
00:00:43
sentences to be used to illustrate
00:00:45
certain syntactic structures in the
00:00:49
subject position the agents machine
00:00:56
Nicole off the state of net conclusion
00:00:57
we hope this study will help readers
00:00:59
realize how important stereotypes are
00:01:01
within our society even though they're
00:01:03
challenged more often nowadays it's
00:01:05
strongly remain thank you so much for
00:01:08
agreeing to participate thank you yeah
00:01:12
thank you so much for inviting me it's
00:01:13
really exciting to be here
00:01:14
so the work I'm going to talk about
00:01:16
today actually just like the work that
00:01:19
Mark has decided is in collaboration
00:01:21
with my graduate students Ilya is she
00:01:23
and in Paris and so my shining point is
00:01:28
that as I'm sure everyone here knows
00:01:29
recent research in a shows that AI
00:01:32
technologies particularly those based on
00:01:34
classifications we furnish the kinds of
00:01:36
biases against women in minority groups
00:01:38
that we sadly find in society overall
00:01:41
why is this well in short it's because
00:01:44
classifiers are trained on corpora that
00:01:46
are produced by humans and humans are
00:01:48
horrible people
00:01:51
all right so in particular it's because
00:01:54
humans tend to hold biases against women
00:01:56
in minority groups and so the production
00:01:58
patterns in down the training corpora
00:02:01
with the bias structure which the
00:02:03
classifier then learns so for example in
00:02:07
her recent review of the kinds of bots
00:02:09
gender bias that we find in corpora that
00:02:12
are commonly used for training Susan Lee
00:02:14
levy lists a number of ways
00:02:16
in which women are disfavored compared
00:02:19
to men in the training corpora so these
00:02:22
include things like naming um ordering
00:02:24
of arguments dyes descriptions metaphor
00:02:28
and the topic of my talk today which is
00:02:30
just the simple presence of expressions
00:02:32
referring to women and men in the text
00:02:34
of the course
00:02:35
so what levy proposes is that a simple
00:02:38
quota system for gender balance gender
00:02:41
balance and training data for machine
00:02:42
learning algorithms may serve to combat
00:02:44
much of the latent bias in text-based
00:02:47
sources of training data and so I you
00:02:50
know as a kind of first pass at this
00:02:52
very difficult problem I think this is
00:02:53
quite reasonable however I think we'd
00:02:55
all agree that how to at the end of the
00:02:58
day effect should effectively
00:02:59
restructure training corpora requires a
00:03:02
deeper understanding of why the corpora
00:03:04
are structured that way to begin with
00:03:05
and so what I'm going to talk to you
00:03:07
about today is um is this problem and in
00:03:11
particular I'm going to give a study of
00:03:13
gender bias specifically kind of overuse
00:03:16
of masculine pronouns in English and
00:03:18
French and so the kind of approach that
00:03:22
I'm going to develop in the talk which
00:03:24
is an approach that we're developing in
00:03:26
my lab in Paris is to combine tools from
00:03:29
formal semantics so I'm a formal
00:03:31
semantics sis with computational psycho
00:03:33
linguistics and feminism so in
00:03:36
particular what we're gonna do is use
00:03:38
mathematics and logic to formalize
00:03:41
hypotheses about linguistic meaning and
00:03:43
gender ideologies and then we're going
00:03:46
to build computational models of the
00:03:48
cognitive mechanisms and representations
00:03:50
that underlie language processing in the
00:03:52
mind of brain and these computational
00:03:54
models will allow us to link our
00:03:57
formalized hypotheses about gender
00:03:59
ideologies for example with predictions
00:04:03
for linguistic production in this case
00:04:04
for pronoun production and the third
00:04:08
kind of ingredient is to really at the
00:04:10
end of the day take seriously the idea
00:04:12
that women currently occupy a
00:04:14
subordinate position in society that
00:04:16
they suffer certain justices and
00:04:18
systemic disadvantages because they are
00:04:20
women and explore the consequences of
00:04:23
this idea for our theories of cognitive
00:04:25
representations
00:04:28
okay so let's look at the patterns of
00:04:31
gender bias in English and French
00:04:32
pronoun use so the starting point of our
00:04:37
talk in our study was a recent study
00:04:40
that came out of Roger levees lab at MIT
00:04:42
voice at all 20:19 where these
00:04:45
researchers were interested in the
00:04:47
relationship between speaker beliefs and
00:04:49
the use of english pronouns
00:04:51
um so their study had sort of two
00:04:53
subparts so the first sub part was
00:04:56
leading an interpretation in fact a
00:04:58
norming study so in their in their
00:05:01
experiment they asked us participants to
00:05:03
read sentences with roll nouns that
00:05:05
varied according to gender stereotypes
00:05:07
and where the the stereotypic allottee
00:05:10
of the noun was was was given by a
00:05:13
previous norming the study Mazursky at
00:05:15
fall 2014 so they were interested in in
00:05:20
sentences with different kind of nouns
00:05:21
for example construction work shirt
00:05:23
worker which is associated with a strong
00:05:25
male stereotype something like Baker
00:05:27
which is associated with a neutral
00:05:29
stereotype and something like manicurist
00:05:31
which is associated with the strong
00:05:33
female stereotype so in their first
00:05:36
experiments they asked participants to
00:05:37
read a sentence like in one so after the
00:05:40
shop on High Street closed for the night
00:05:42
a baker stick to tidy up and then they
00:05:45
asked participants how likely is it that
00:05:47
the individual described in the sentence
00:05:49
is a woman and so this first experiment
00:05:51
allowed voice at all to develop a
00:05:53
measure of of the the belief of
00:05:57
participants and the social gender of
00:05:59
the referent of the noun phrase like a
00:06:01
baker after they read a sentence so they
00:06:07
had this measure of belief and then the
00:06:09
second part of the experiment was a
00:06:10
production experiment in particular a
00:06:12
close continuation study and so they had
00:06:16
participants read one of the sentences
00:06:18
from the interpretation study and then
00:06:20
wrote another half sentence and then
00:06:23
asked them to complete the sentence the
00:06:25
second sentence in whatever way they
00:06:26
want okay
00:06:28
so they'd read after the shop on high
00:06:30
she closed for the night a baker state
00:06:32
decided yeah
00:06:32
before the baker takes up a trash and
00:06:34
then participants just wrote whatever
00:06:36
they want so sometimes yeah they just
00:06:39
wrote all sorts of things but in the
00:06:40
vast
00:06:41
majority of cases they used a pronoun
00:06:43
that referred to a baker for example and
00:06:46
yeah and so that this is what voice at
00:06:49
all studied so the results of their
00:06:53
study they are shown on the slide so on
00:06:56
the x-axis here you see the results of
00:06:58
their interpretation study so I'm the
00:07:01
degree of belief that in that the
00:07:04
reference is a woman so so zero over
00:07:08
here
00:07:08
means that participants are certain that
00:07:11
the reference man and one is that
00:07:14
they're certain that we're talking about
00:07:15
a woman and on the y-axis here you see
00:07:18
the proportion of use of pronouns and so
00:07:22
and this yeah so proportion of use of
00:07:25
pronouns of he she and thing the first
00:07:29
thing you can observe from the scrap is
00:07:31
the participants in voice it all study
00:07:33
they're not big users of the singular
00:07:35
they
00:07:36
how do anyone actually use today in
00:07:38
particular they mostly just I'm very
00:07:40
between he and she and and the other
00:07:44
thing you can notice from this graph is
00:07:46
that participants in the study they only
00:07:48
use she when they are either certain or
00:07:50
almost certain that the reference is
00:07:52
female
00:07:52
okay so particular there's a zone kind
00:07:56
of around here where the participants
00:07:59
believe it's more likely than not that
00:08:01
the reference is female and yet they
00:08:03
almost exclusively use heat so this is
00:08:07
what we're going to talk about in terms
00:08:08
as a as a gender bias and is overuse of
00:08:11
masculine pronouns make some cases where
00:08:13
the use of a pronoun
00:08:15
in this case a maximum program doesn't
00:08:17
align with the beliefs in the social
00:08:19
gender of the speaker alright so this is
00:08:23
a this is a curious pattern we want to
00:08:27
know why where it comes from
00:08:29
why we find it and you know there could
00:08:32
be a number of things going on we want
00:08:33
to know is it you know because of a
00:08:35
particular the particulars semantics of
00:08:37
the English pronoun system does it have
00:08:39
to do with gender ideologies
00:08:42
what's going on so something that will
00:08:45
help us decide between the different
00:08:47
possible hypotheses is to know how
00:08:49
general this kind of overuse of
00:08:51
masculine pronoun pattern is
00:08:54
exclusively in English or do we find it
00:08:56
in other languages so we were interested
00:08:59
in this question so we decided to look
00:09:01
at French now French is an interesting
00:09:04
language for these kinds of things
00:09:05
because unlike English it's a
00:09:07
grammatical gender language and what
00:09:09
that means is that French grammar sorts
00:09:10
all nouns into two classes masculine or
00:09:13
feminine which then go on to determine
00:09:15
patterns of agreement with other
00:09:17
expressions in the sentence so the first
00:09:20
thing to note is that dik-dik uses of
00:09:22
most if not all gender marked
00:09:24
expressions referring to human nouns
00:09:27
I strongly create male or female
00:09:29
interpretations okay so so if I take a
00:09:32
neutral stereotype man it's journalist
00:09:35
and I say click on neutral Minister yeah
00:09:38
um this very strongly me seems to
00:09:42
communicate look the male journalist is
00:09:44
there likewise if I say me young latin
00:09:46
men you see though you really want to
00:09:49
interpret that you're talking about a
00:09:51
female journalist and similarly for the
00:09:53
pronouns since bread french is the
00:09:59
grammatical gender language it has a
00:10:01
robust agreement system and sometimes
00:10:04
agreement can disrupt the alignment
00:10:06
between grammatical and social gender
00:10:08
sometimes actually quite radically as
00:10:10
you see in the sentence here so this is
00:10:12
a sentence that comes from a documentary
00:10:15
about periods and it says so vision is
00:10:19
all dolled up with heels
00:10:20
yeah the Johnnies only like it actually
00:10:23
delivers so there are people who have
00:10:25
endometriosis there people have horribly
00:10:27
painful periods and in this case so the
00:10:31
noun Jean is masculine and so the
00:10:33
program that follows it is either the
00:10:36
masculine as well despite the fact that
00:10:38
virtually everyone we're talking about
00:10:40
here is a woman alright so in order to
00:10:47
see how grammatical gender interacts
00:10:49
with pronoun use we essentially
00:10:52
replicated the voice at all studies and
00:10:55
so we first did some norming studies so
00:10:58
we had 36 common gender nails that were
00:11:01
balanced for stereotype based on the
00:11:04
Mazursky score metric and so crucially
00:11:07
we looked at only nouns that in French
00:11:10
were from the the common gender pattern
00:11:13
so these are nouns that do not vary the
00:11:16
form of the noun depending on the
00:11:18
grammatical gender so so something like
00:11:20
humanist journalists you say visual -
00:11:22
last one - you don't change the form of
00:11:24
the noun so we did not look at nouns
00:11:26
that change the form of the noun like
00:11:28
for example no Boulanger yeah Boulanger
00:11:30
like Baker so we actually distinguished
00:11:36
we did two studies so the first one
00:11:38
where you could was a kind of normal
00:11:40
case where you see the grammatical
00:11:42
gender marking so we had purchased
00:11:44
asleep read a sentence like six okay and
00:11:49
then Luna and then you know a variety of
00:11:51
different so when they're calling the
00:11:59
diplomat secretary journalist
00:12:01
always calls with a hidden number and
00:12:05
then we also because we're interested in
00:12:06
what the contribution of the grammatical
00:12:08
gender is to two interpretations of
00:12:12
social gender and pronoun use we also
00:12:14
were interested in cases where for
00:12:17
morphological reasons the grammatical
00:12:19
gender is neutralized and so this
00:12:22
happens for example in with noun phrases
00:12:24
with pre the pre verbal vowel initial
00:12:27
comment in there adjectives so things
00:12:29
like 7-pull pitifully Nicki they're
00:12:32
eager to grab a spell table that's a
00:12:36
good doesn't it
00:12:37
after - the only male machine right so
00:12:39
here because of the phonological form of
00:12:42
the adjective you can't tell whether
00:12:44
it's masculine or feminine and so this
00:12:48
gives us this will allow us to get a
00:12:50
measure of the speaker's beliefs in the
00:12:53
gender the social gender of the referent
00:12:55
after they read these sentences so
00:12:58
here's the results of the interpretation
00:13:01
study so this is from the neutralization
00:13:04
so the this case here where you can't
00:13:06
tell if there's gender on the x axis
00:13:09
here you see the Mazursky stereotype
00:13:11
score where 0 is the associated with the
00:13:15
strongest male stereotype and one is
00:13:18
with the strong female stereotype and on
00:13:22
the y-axis you see the belief the belief
00:13:25
or the likelihood according to
00:13:28
participants that the reference is
00:13:29
female and so perhaps unsurprisingly
00:13:32
what you'd find is that with the more
00:13:34
stereotypically male now participants
00:13:38
are more likely to think that we're
00:13:40
talking about a man and then it goes so
00:13:47
particular what is interesting is that
00:13:49
in our results for the neutralization
00:13:51
the speaker's belief actually basically
00:13:53
reproduces the mazurski score and so
00:13:57
this is um you know understandable and
00:13:59
yeah that's understandable because here
00:14:01
you don't have grammatical gender so
00:14:03
you're just using stereotypes to
00:14:05
influence your movie so so that's for
00:14:10
the new toy section cases and then if we
00:14:11
add dramatical gender what happens
00:14:13
essentially is that all it does is it
00:14:17
for the masculine decreases the belief
00:14:19
or increases the belief that you're
00:14:21
talking about a man and for feminine
00:14:23
increases the belief that you're talking
00:14:24
about a work so we have the stereotype
00:14:27
of stereotypes here and then for
00:14:30
masculine right you just it just pushes
00:14:33
you into the to believe anymore that
00:14:36
you're talking about a guy but you still
00:14:38
find a stereotype effect and the
00:14:40
feminine just pushes you up Lily makes
00:14:43
you believe that it's more likely that
00:14:45
you're talking about a woman but again
00:14:47
even in the case of the feminine you
00:14:48
still find a stereotype effect so that's
00:14:52
I'm the results of the interpretation
00:14:54
study and then we did the continuation
00:14:57
study where we added a second half of
00:15:00
the sentence and asked them to complete
00:15:02
it however they wished so again
00:15:05
sometimes we got some sort of crazy
00:15:06
things but most of the time people use
00:15:09
the promo and so we see the results here
00:15:13
like this so this is for the
00:15:15
neutralization case right so we don't
00:15:17
have grammatical gender so here on the
00:15:21
slide you see the mean value in the
00:15:24
neutralization experiment going from
00:15:27
0.25 and on the y-axis
00:15:32
is essentially the proportion of uses of
00:15:34
the feminine pronoun and so I'm the
00:15:38
first thing that we observe is that we
00:15:40
find the same pattern in French as
00:15:43
what's it all found in English where and
00:15:47
in particular what we find is that with
00:15:52
females stereotyped professions heaps of
00:15:55
you know matches the participants belief
00:15:58
in the gender of the room so if you're
00:16:01
over in this corner of the graph here in
00:16:03
unannounced with very high strong female
00:16:07
stereotypes then participants use of so
00:16:11
if participants believe that it's 75%
00:16:15
likely that we're talking about four
00:16:17
women then they use about 75% these
00:16:21
about 75% in fact they actually use
00:16:25
right so on this side here the use of
00:16:29
the pronoun matches the belief but when
00:16:33
you get down over here into the strongly
00:16:36
may have stereotyped professions then we
00:16:39
find a disalignment
00:16:41
in particular so you think it's 75%
00:16:45
likely that we're talking about a man
00:16:47
you're going to use either almost 100%
00:16:49
of the time so um so it's so in
00:16:54
particular we find a case of overuse of
00:16:57
the pronoun in in with now that have
00:17:02
strong math masculine stereotypes so
00:17:06
this is the pattern that we're going to
00:17:07
mostly talk about for the rest of the
00:17:09
talk the neutralization case but in case
00:17:13
you're interested in what happens with
00:17:15
the grammatical gender and agreement so
00:17:17
long story short agreement does seem to
00:17:20
play a large role in ensuring that the
00:17:22
pronoun that you use will kind of match
00:17:25
the grammatical gender of the noun
00:17:27
phrase that it's referring to however
00:17:30
even so we still find in a symmetry and
00:17:32
asymmetry that favors the masculine
00:17:36
particular we find in our study that
00:17:39
agreement mismatches right so cases
00:17:41
where the
00:17:42
oh gender of the pronoun doesn't match
00:17:44
the grammatical gender of the noun
00:17:45
phrase they're more likely if the noun
00:17:48
phrase is feminine so our participants
00:17:50
if they read something that logical
00:17:53
match so feminine Marc diplomat will go
00:17:55
the block they're more likely to - if
00:17:57
they're going to - not to break
00:17:59
agreement they're more likely to say
00:18:02
that if participants say good spiel that
00:18:06
[Music]
00:18:15
okay so we have this pattern of of
00:18:18
gender bias and now we want to know why
00:18:21
do we have it and you know these things
00:18:23
are very complicated so what I'm going
00:18:26
to do is is use the tools of formal
00:18:30
semantics and computational cycle
00:18:31
Mystics to try to formalize some some
00:18:35
ideas that we might some hypotheses that
00:18:36
we might make but why is that
00:18:38
and then in and the formalization will
00:18:43
allow us to precisely assess what the
00:18:47
predictions of the different hypotheses
00:18:49
are which we can then compare to the
00:18:51
statistical models based on our actual
00:18:53
data so we'll look at two different
00:18:59
kinds of hypotheses so one that comes
00:19:02
mobile more at least currently in France
00:19:05
is advocated more by people coming from
00:19:08
the sort of conservative political
00:19:10
spectrum and one by the other hypothesis
00:19:14
that has been advocated for primarily by
00:19:17
people in the feminist tradition the
00:19:20
first hypothesis is to say okay you know
00:19:24
we have these kind of over uses these
00:19:26
like extra uses of the in you know
00:19:29
referring to to someone that this the
00:19:32
participant you know presumably thinks
00:19:34
is probably a woman
00:19:35
well how can we have this it's a well
00:19:38
actually it's because um the semantics
00:19:41
of it is not actually a strictly male
00:19:44
but in fact it is a kind of default
00:19:47
gender-neutral permit so this idea has a
00:19:51
long-standing um you know in both the
00:19:53
structuralist
00:19:55
and adopted many people and so so it's
00:20:00
basically the idea that that even has a
00:20:01
gender-neutral smell
00:20:03
now it's very easy to formalize this
00:20:06
idea in formal semantics and so
00:20:10
following our work for example by Robin
00:20:12
Cooper or honey Kratzer we all assume
00:20:14
that pronounce what they do is they
00:20:15
trigger presuppositions concerning the
00:20:18
social gender of their record and so we
00:20:21
can write this in formal semantics
00:20:23
notations as in 10 a and 10 B
00:20:26
essentially what the sense is that the
00:20:28
pronoun and it's only going to be the
00:20:31
the use of a pronoun L is only going to
00:20:34
be felicitous or or just fine if the
00:20:38
reference of the pronoun has this
00:20:41
particular socially defined property of
00:20:44
being female now even on the other hand
00:20:48
so the gender-neutral analysis of is we
00:20:51
would say well it can be defined
00:20:53
if the pronoun is male or or if the
00:21:01
reference is people all right so that so
00:21:05
it's generally neutral semantics now of
00:21:09
course you know someone who wants to
00:21:10
come and tell me or tell anyone that ian
00:21:13
has gender-neutral semantics faces an
00:21:15
immediate puzzle and that's what we saw
00:21:17
before why is it that well if it is
00:21:20
gender-neutral why is it that if you say
00:21:22
ah yeah you know this is extremely
00:21:24
bizarre if the person you're talking
00:21:27
about doesn't identify our person
00:21:29
doesn't matter so the way that these
00:21:32
kind of cases have been treated in the
00:21:33
literature is to say is the following
00:21:36
it's just say why do we get these male
00:21:39
interpretations of even if it's actually
00:21:41
gender-neutral well it's because the
00:21:43
male interpretations arise as a result
00:21:46
of an implicature so since if you
00:21:49
compare the denotations at 10 a and 10 B
00:21:52
since today is more informative than to
00:21:54
me for communicating female social
00:21:56
gender the speakers YouTube in is going
00:21:59
to allow the listener to conclude that
00:22:00
the reference is a man because if they
00:22:02
really meant to communicate that they
00:22:04
were a woman they wouldn't used that
00:22:07
so yes so the nail implicature so the
00:22:11
male inference is a result of a great
00:22:13
scene implicature fortunately for us in
00:22:16
computational cycles we have a number of
00:22:18
computational models aggression
00:22:20
reasoning the one that I don't have time
00:22:23
to go through them here the one that we
00:22:24
use is called the rational speech act
00:22:27
framework that maybe some people here
00:22:29
are familiar so we have our promo
00:22:32
semantic analysis now we can put it in
00:22:35
the rational speech act a computational
00:22:38
model of gray suit reasoning and it will
00:22:40
give us predictions for the use of the
00:22:42
pronet and the RS units like here so
00:22:47
this is the statistical model that we
00:22:50
saw earlier and this is the predictions
00:22:53
of the gender-neutral model as you can
00:22:55
see when you get to the stereotype point
00:23:08
there's a really big difference and I'm
00:23:12
really at the end of the day the reason
00:23:14
why there's this big difference is
00:23:16
because as I described to you earlier in
00:23:19
our data our participants are kind of
00:23:21
treating the stereotype of people
00:23:25
stereotype noun and the male stereotype
00:23:27
nouns differently so the idea is that if
00:23:31
you if you want to match program use
00:23:33
with male oh man stereotypes you have to
00:23:36
say that speakers must rarely draw the
00:23:37
male implicature right so if you want to
00:23:39
explain why they're using so much eel
00:23:41
over here you have to think it's because
00:23:43
it is really gender neutral and they're
00:23:45
not drawing these these male
00:23:46
implicatures so much so they can even
00:23:48
over here but then then when you get out
00:23:51
to the you know part of the graph where
00:23:54
people are actually matching their
00:23:56
belief they don't have to start saying
00:23:57
oh I know it's because they're drawing
00:23:59
and click you know male implicatures all
00:24:01
the time and that's not really how we
00:24:04
think things like informative 'ti and
00:24:06
gretchie and reasoning works so I mean
00:24:09
we don't think this is a great model for
00:24:12
our data so I'm the second hypothesis
00:24:16
I'm
00:24:16
to talk about in the last couple minutes
00:24:18
is is is one in which we say so instead
00:24:22
of saying you know native people are
00:24:24
using work either because this is kind
00:24:27
of default pronoun maybe actually um
00:24:31
people are using either because there's
00:24:33
a because of the relationship between in
00:24:35
and the social category mail but
00:24:38
actually in our ideologies in our
00:24:40
semantic model there's a certain special
00:24:43
struck there's a certain structure to
00:24:45
our semantic models and our ideological
00:24:46
models such that it is mail is actually
00:24:49
a default gender category and so this is
00:24:53
a proposal that is featured very
00:24:54
prominently in feminist thinking both in
00:24:58
France and elsewhere but of course you
00:25:02
know the idea of mail is a kind of
00:25:04
default category social category it's
00:25:09
very abstract and so we need to do I'm
00:25:11
still work to formalize this idea to to
00:25:14
bring it in line to test its predictions
00:25:16
with our pronouns data so that's what
00:25:20
we'll do um so the idea is that we're
00:25:23
going to take our semantic model and
00:25:25
we're going to enrich it with dominance
00:25:28
relations and so in particular in the
00:25:31
model will say that individuals in the
00:25:32
extension of this social of a social
00:25:36
properties male and female will be in a
00:25:39
dominance relation such that the middle
00:25:42
are the extension the individuals in the
00:25:44
extension of male will dominate the
00:25:47
individuals in the female and so we'll
00:25:50
just be code this as dominate relations
00:25:53
here alright so we have these these
00:25:58
ideological representations of dominance
00:26:00
and so how what is the relation between
00:26:02
them and and how we use pronouns and so
00:26:06
we think that the relation between
00:26:09
ideological representations of dominance
00:26:11
and language use passes through
00:26:13
similarity so we know from psychological
00:26:19
studies of similarity that have nothing
00:26:23
to do with the gender domain that
00:26:25
dominance effects similarity judgments
00:26:28
so for example
00:26:30
now classic study by severe ski in 1977
00:26:32
um he asked participants to compare
00:26:36
different countries and he found that
00:26:39
participants are more likely to judge
00:26:40
the less dominant left powerful country
00:26:42
to be similar to the Morgul dominant
00:26:45
powerful country than vice-versa
00:26:48
right so if you ask people oh it's Korea
00:26:51
similar North Korea similar to Red China
00:26:53
these are examples from the seventies
00:26:57
Red China they'll be like yeah sure but
00:27:01
if you ask them is Red China similar to
00:27:03
North Korea then they're more likely to
00:27:05
say no
00:27:05
similarly you ask them so is next go
00:27:08
similar to the United States not so sure
00:27:17
so yeah so we're gonna encode these non
00:27:23
symmetries between it's between powerful
00:27:27
and less powerful individuals as in
00:27:32
terms of similarity relations in our
00:27:34
semantic models and they're subject to
00:27:36
the constraint in 50 so if a a one
00:27:39
dominates hu then a one is not going to
00:27:42
be viewed similar in the model to 80 now
00:27:47
crucially um so we have a similarity
00:27:50
relations we have a constraint on them
00:27:51
that reflects ideological perceptions of
00:27:54
dominant and um
00:27:56
furthermore we're going to say that
00:27:57
what's gonna happen in our model is that
00:27:59
similarity relations are gonna change
00:28:01
depending on stereotype classes so this
00:28:03
is an idea that has been explored a lot
00:28:06
in feminism in France and it's the idea
00:28:10
that you know the more male dominated
00:28:12
the profession is the war women in it
00:28:14
are going to be considered
00:28:16
I'm similar to men so if you have a very
00:28:18
male-dominated profession like for
00:28:19
example AI technologists then you're
00:28:23
good then the women in there are going
00:28:25
to be good to be kind of grouped
00:28:26
together with the men but if you have a
00:28:28
very female dominated area the men are
00:28:31
not gonna B's are considered to be
00:28:32
similar to to the way and so our model
00:28:38
will have the proportion of women who
00:28:41
the speaker uses similar
00:28:42
decrease in this case literally I don't
00:28:45
they believe that the reference is a
00:28:46
woman increases and finally on the final
00:28:51
bit of the relation between dominance
00:28:52
relations and language use comes through
00:28:56
the fact that these similarity judgments
00:28:58
right that are affected by dominance
00:29:00
relations we also know the effects of
00:29:02
the meanings of words and a linguistic
00:29:05
categorization so um well we work in on
00:29:08
big language we can construct what's
00:29:11
called a tall order kind of loose
00:29:13
interpretations of these these social
00:29:17
gender properties in the following way
00:29:24
so we'll say that an individual a one is
00:29:26
in the tolerant extension of man just in
00:29:31
case it is similar a one is similar to
00:29:34
some individual that is in the kind of
00:29:36
core basic extension and similarly for
00:29:40
female and then you have the semantics
00:29:44
of the pronouns defined in terms of
00:29:47
these tolerant interpretations and so
00:29:55
this is the way in which we're going to
00:29:58
formalize in our semantic model and in
00:30:01
the computational model this idea the
00:30:03
feminist idea that comes with even since
00:30:06
as of male as being the kind of default
00:30:09
gender category okay so the idea which
00:30:13
is described by a penny Eckert and
00:30:14
selling college name in their language
00:30:16
and gender textbook that Marcus put up
00:30:18
on this on the slide earlier so they
00:30:21
they observed that you know while TVs
00:30:22
and behaviors labeled as mingle are
00:30:24
treated as appropriate for females as
00:30:27
well as masked those label this female
00:30:29
are treated as appropriate only for
00:30:30
Phoenix one way of looking at this is
00:30:32
that female activities and behaviors
00:30:34
emerges marked as a reserved for a
00:30:37
special subset of the population
00:30:38
well male activities and behaviors
00:30:40
version on marker normal
00:30:42
this in turn contributes to the end of
00:30:44
centric and Andrew centric male centric
00:30:46
view of gender so this is a
00:30:48
formalization interpoma semantic models
00:30:51
on this idea until it so the structure
00:30:54
of the models with this gives us
00:30:55
is the fact that in the model women can
00:30:59
be considered loosely or tolerantly men
00:31:01
but men cannot be considered loosely or
00:31:03
tolerantly women and there's a theorem
00:31:05
to this effect so now we have a model
00:31:08
and we look at what predictions it makes
00:31:19
and so if I had to sum up why exactly
00:31:23
this amount this works it's because so
00:31:27
it's strongly male coded professions
00:31:29
you're gonna have many women so these
00:31:35
can be used to refer to them and so this
00:31:37
over in this corner here you're gonna
00:31:38
get a kind of what we were calling an
00:31:40
overuse of P but then it's strongly
00:31:43
strongly female coded professions you'll
00:31:46
have fewer or to know what extension of
00:31:50
male and so by the time you get over
00:31:53
here then they use it in it's going to
00:31:55
match your belief all right so you
00:31:59
conclude although I think this kind of
00:32:01
brute force quota idea as a fixed to
00:32:05
treating corpora you know it's probably
00:32:07
better than nothing fully eliminating
00:32:09
gender bias requires a deeper
00:32:11
understanding of where and why gender
00:32:12
bias are cursed so what we found in our
00:32:15
studies and which is the famous
00:32:17
voiceover English the overuse of
00:32:20
masculine prize it was actually limited
00:32:22
to stereotypically male professions okay
00:32:25
so if you have a training corpus and you
00:32:26
want to eliminate go in and eliminate
00:32:28
male pronouns you should be doing this
00:32:30
in the stereotypically male profession
00:32:33
part of the corpus but not in the
00:32:36
stereotypical and I've also argued that
00:32:42
we can formalize hypotheses be they
00:32:45
coming from feminists of the more
00:32:47
conservative corners of the political
00:32:49
spectrum using formal semantics and
00:32:51
assess what predictions they make for
00:32:53
empirical data using computational
00:32:56
linguistic models so this kind of gives
00:32:58
another sorts of argumentation between
00:33:00
these different ideas about language and
00:33:02
meaning and in this case and we argue
00:33:05
that models of encode feminist ideas
00:33:07
about male as
00:33:08
default gender category make better
00:33:10
predictions for English and French
00:33:11
pronunciation
00:33:15
[Applause]
