►
From YouTube: NLP in Clojure Session 1
Description
Following a few recent suggestions, we organized a study meeting about NLP, possibly a first on a series.
This event is related to the upcoming ds4clj course – it is part of our exploration of topics that may later become part of the course.
Moderated: Kira McLean
The session began with an intro by Dimid Duchovny about Linguistics and Natural Language Processing.
Then, Daniel Slutsky presented the use of the spaCy library, first from Python, and then from Clojure through Libpython-clj.
Towards the end, Carsten Behring added some comments about the broader picture of interop and machine learning in Clojure.
Summary page at Clojureverse: https://clojureverse.org/t/nlp-in-clojure-session-1-summary-recording/
Zulip chat of the session:
https://clojurians.zulipchat.com/#narrow/stream/330924-ds4clj/topic/nlp.20session.201.2C.202022-09-10
A
A
So
for
now
this
is
kind
of
just
a
one-off
meeting
that
came
out
of
a
desire
from
some
people
in
the
community
to
learn
about
this
stuff,
but
probably
we'll
have
more.
So
we
have
many
ongoing
study
groups
in
the
cyclo
community
for
certain
topics
like
data
visualization
or
data
processing
whatever,
and
so
it
seems
likely
that
there's
enough
interest
to
start
a
natural
language
processing
one
as
well,
but
we'll
see
how
it
goes.
A
There's
also
a
course
coming
up
that
I
wanted
to
mention.
So
there's
been
some
interest
in
the
community
and
some
efforts
from
a
few
of
us
to
put
together
a
sort
of
more
cohesive
data,
science
enclosure
course,
and
that
would
probably
take
the
form
of
like
a
monthly,
roughly
kind
of
half
a
day
session
where
we
would
get
together
and
someone
would
cover
a
topic
and
then
we'd
have
some
time
to
explore
together
and
learn
on
our
own.
A
I
guess
semi-related
to
that
is
also
another
project
that
I'm
working
on
is
a
closure
data
cookbook,
which
is
hopefully
going
to
be
kind
of
like
a
companion
guide
to
that
kind
of
thing.
So
the
idea
is
for
it
to
be
like
a
sort
of
like
a
book
but
kind
of
more
interactive.
More,
I
don't
know
like
a
living
kind
of
website
type
thing,
we're
not
too
sure
exactly
what
the
published
like,
what
what
tools
we'll
use
to
publish
it.
A
But
the
goal
is
that
to
be
like
a
sort
of
living
example,
collection
of
examples
for
how
to
use
a
bunch
of
these
newer
closure
libraries
to
do
data
science
tasks
in
closure
and
yeah.
I
guess
kind
of
the
high
level
like
vision
or
direction
or
whatever
that
the
cyclist
community
is
hoping
to
move
in,
is
basically
just
opening
up
to
other
communities.
So
like
right
now,
sorry,
I've
got
two
dogs
here,
so
they're
a
little
bit
chaotic
right
now
also
they're
very
excited
about
this
topic
yeah.
A
A
That
makes
it,
I
think,
a
really
viable
stack
if
you
will
for
that
kind
of
work,
and
so
that's
kind
of
like
the
big
idea
that
we're
kind
of
working
on
is
like
trying
to
build
tools
and
resources
and
examples
and
courses
and
whatever,
so
that
people
can
see
that
it's
also
a
valuable
and
useful
set
of
tools
for
that
kind
of
work,
and
let
me
just
double
check.
I
had
a
little
couple
things
notes
of
things
that
I
wanted
to
cover.
A
I
think
yeah,
so
the
goal
for
yeah
so
for
I
guess
just
also
to
mention
this
meeting.
It's
obviously
about
closure.
This
is
a
cyclist
community,
but
we're
open
to
anyone
like
in
general.
So
it's
not
just
we're
not
assuming
very
much
background
and
specifically
enclosure
at
all,
but
on
the
other
hand,
if
you
do
know
closure-
and
you
don't
know
anything
about
natural
language
processing,
that's
fine
too.
We're
kind
of
hoping
to
bridge
those
worlds,
and
so
that's
kind
of
the
goal.
A
With
this
study
group,
the
other
study
groups,
the
quickbook
and
the
course,
and
all
of
it
is
to
like
bring
together
closure
developers
and
people
who
are
building
these
libraries
and
tools
with
people
who
primarily
just
work
with
data
in
a
sort
of
tool,
agnostic
way
and
kind
of
like
merge
them.
So
today
in
particular,
we
have
a
special
guest
damid
danid.
Sorry,
if
I'm
mispronouncing,
that
damid,
who
is
mostly
a
data
scientist
working
with
python
and
is
going
to
teach
us
a
little
bit
about
natural
language
processing.
A
So
we'll
spend
the
first
kind
of
20
30
minutes
on
that,
and
then
we'll
go
over
how
to
do
some
of
that
stuff
or
how
basically,
how
to
interact
with
these
python
libraries
enclosure,
which
will
be
led
by
daniel
who's,
the
other
sort
of
host
here
and
then
at
the
end,
we're
hoping
to
save
the
last
kind
of
half
of
the
meeting
for
basically
like
a
kind
of
interactive
like
collaborative
work
session.
So
it's
kind
of
open
kind
of
up
in
the
area.
A
A
B
C
Thank
you
kira
and
danny,
so
I'll,
just
introduce
myself
briefly.
Well,
I'm
a
data
scientist
I
work
mostly
in
llp
problems.
My
background
is
computer
science
and
linguistics,
and
for
the
last
year,
or
so
I
had
the
honor
to
be
danny's
colleague,
we
actually
first
met
when
we
were
danny,
organized
some
sessions
about
functional
programming,
and
it
was
mostly
about
closure,
and
I
was
coming
from
a
haskell
background,
so
we
found
or
tried
at
least
to
find
the
common
ground
yeah.
C
C
C
C
We
we
see
it,
we
find
it
in
our
everyday
lives
very
common.
Nowadays,
virtual
assistants,
google
search
translations.
You
can
practically
find
it
everywhere
and
just
a
note
about
terminology,
there
is
another
kind
of
nlp
which
stands
for
neurolinguistic
programming,
so
it
was
once
approached
by
someone
in
linkedin
who
was
an
nlp
master.
Just
he
won't
confuse
the
two
okay.
B
D
C
C
C
C
For
example,
we
have
our
language,
which
is
basically,
as
the
shishu
proposed,
is
some
coupling
of
a
sign
and
a
concept,
and
finally,
as
people
with
a
background
in
computer
science,
you
probably
appreciate
that
language
or
language
is
recursive
example.
Danny
said
the
kira
said
that
robert
cats,
that
they
meet
said
and
so
on,
it
could
go
at
infinite.
C
When
we
speak
about
human
languages,
we're
talking
about
different
families
of
languages,
for
example,
the
most
known
one
might
be
the
indo-european
languages.
You
also
have
the
austrian
ancient
languages,
the
and
the
other
tabs
launches.
I
don't
want
to
go
in
too
much,
and
here
you
can
see
the
tree
and
as
computer
science,
people
you'll
appreciate
this
data
structure.
I
believe
yes,
this,
for
example,
is
for
the
indo-european
languages
and
you
can
see,
for
example,
english
and
languages
from
the
indian
subcontinent,
for
example,
hindi
and
urdo.
C
Okay,
so
let's
talk
about
a
bit
about
linguistics
and
what
which
subfields
do
we
have
in
listings?
So,
let's
start
from
maybe
the
most
bottom
level
and
we'll
go
our
way
out.
So,
first
of
all,
when
you're
talking
about
spoken
language,
we
have
talking
about
sounds
which
are
called
phones,
and
the
study
of
such
phones
is
called
phonetics
here,
for
example,
you
can
see
different
sounds
and
their
articulation.
C
Some,
for
example,
p
or
b
are
bilabel,
which
means
they're
pronounced
with
both
ellipse,
and
there
are
other
sounds
which
are
pronounced
even
further.
This
is
called
the
articulation
place.
We
also
have
distinctions
between
consonants
and
vowels
and
fricatives
and
which,
like
or
stops
like
putin,
but
I
won't
go
into
all
the
details
now.
C
Next,
we
can
group
different,
sounds
into
categories
called
phonemes
and
basically,
when
we're
we're
talking
about
a
language
where
usually
we
usually
don't
want
to
be
very,
very
nuanced,
but
but
instead
talk
about
which
phonemes
it
has
okay.
So
just
a
word
about
this.
This
is
in
the
ipa,
the
national
phonetic
alphabet,
which
is
just
an
alphabet,
but
not
of
written
language
but
spoken
language.
C
The
basic
unit
is
not
a
letter,
but
the
sound.
So
these
are
the
consonants,
and
here
are
the
vowels
and,
as
we
say,
we're
talking
about
specific.
So
the
ipa
is
international,
it's
universal,
more
or
less,
and
when
we're
talking
about
specific
language,
you're
more
interested
in
not
informatics,
but
rather
in
phonology,
which
means
which
phonemes
are
useless
in
language,
and
the
reason
is
that
many
people
have
their
own
way
of
pronouncing
things,
and
these
differences
might
not
matter
so
much.
They
don't
affect
our
understanding.
C
Let
me
elaborate,
for
example,
we
have
a
concept
that
is
called
a
minimal
pair,
which
is
the
minimal
change
known,
which
affects
the
meaning
of
the
word.
So,
for
example,
if
you
take
the
first
pair
pen
and
pin
in
english,
these
are
two
different
words
but,
for
example,
in
a
language
such
as
arabic,.
C
C
Okay,
okay,
so
the
next
subfield
is
morphology
and
mythology.
Morphology
deals
with
how
how
diff
different
subunits
make
a
word.
So,
for
example,
let's
talk
about
independently
right,
we
have
this
prefix
in
the
main
verb
depend
an
affix
and
and
another
efficacy.
Okay,
so,
and
each
such
basic
part
is
called
the
morpheme,
and
when
we
are
talking
about
morphology,
we
basically
are
talking
about
two
kinds
of
morphemes,
one
of
them
and
the
first
one
is
called
inflection,
and
these
are
used.
C
For
example,
in
the
context
of
works,
we
can
select
the
verb
in
time
in
in
in
person.
Maybe
in
aspect
depends
on
the
language
in
mode
and
the
other
kind
of
morph,
and
this
morph
basically
doesn't
change
the
meaning
very
much
in
the
sense
that
it's
still
a
verb,
but
we
have
another
kind
of
morphine,
which
is
a
called
interventional
morpheme,
and
it
basically
turns
one
part
of
speech
into
another
one,
possibly,
for
example,
from
the
verb.
C
C
C
For
example,
the
very
small
boy
is
to
gather
a
known
phrase
and
kiss
the
platypus
is
a
verb
phrase,
because
its
head
here
is
the
word
kissed,
which
is
a
verb
and
in
the
noun
phrase,
the
very
small
boy.
The
head
is
basically
the
boy
which
is
a
noun
this
kind.
This
kind
of
formalism
is
called
a
constituency
grammar
and
we
won't
talk
about
it
today.
A
lot
we'll
talk
about
some
other
kind
of
formalism,
which
we'll
see
in
a
second.
C
C
So
let's
talk
a
bit
about
why
nlp
is
a
difficult
task,
so
is
it
first
unlike
languages
such
as
english,
where
you
can
distinguish
between
words
using
white
space
in
some
languages?
It's
not
the
case
exactly,
for
example,
japanese,
here
and
now.
Another
example
comes
from
what
is
called
morphologically
rich
languages.
For
example
hebrew
you
can
see
here
the
word
the
the
letters
bet
sadi.
C
C
C
And
finally,
an
example,
courtesy
of
nam
homskey,
which
says
he
wanted
to
give
an
example
of
something
that
makes
perfect
sense,
syntax
wise,
but
doesn't
mean
anything
semantics
wise
right,
so
colorist
green
ideas
slip
furiously
it's
a
perfectly
valid
sentence,
but
what
does
it
mean
right?
So
we
can
actually,
our
language
faculty
allows
us
to
create
such
sentences,
which
don't
have
a
meaning
but
then
nonetheless
very
valid
sentences
in
our
grammar.
C
Okay,
let's
talk
about
another
interesting
type
of
ambiguity,
see
when
we
have
the
sentence.
Sherlock's,
not
the
man
using
binoculars
actually
allows
two
meanings
right:
either
sherlock
is
the
one
who
is
holding
the
binoculars
and
looking
at
another
man
or
sherlock
is
looking
at
another
man,
and
this
other
man
is
using
binoculars.
C
C
Instead,
we
have
this
concept
of
dependency.
We
start,
we
start
with
a
concept
called
the
root
which
you
might
be
familiar
with.
If
you've
dealt
with
computer
science,
trees
right,
which
ever
wrote
and
from
this
root,
we
have
edges
the
different
words
of
the
sentence
right
which
and
this
and
when
you
have
an
edge
from
one
word
to
another,
it
means
the
the
the
other
word
is
dependent
of
the
of
its
parent.
Okay.
C
Here
we
can
see
a
few
different
grammatical
relations.
You
know
this
in
the
nominal
subject.
We
also
have
a
direct
object,
indirect
objects,
and
so
on
this,
this
formalism
comes
from
a
framework
called
universal
dependencies,
which
was
a
basically
a
very
large
effort
between
academia
and
the
industry
to
basically
define
such
standard,
which
would
be
used
by
everyone
and
they've
done
a
great
job.
C
C
Yes,
I
just
wanted
to
say
that
if
you
know
spanish,
that
you
know
that
in
spanish
we
have
the
well,
we
have
it
in
english,
but
only
a
bit
in
spanish.
There
is
a
big
difference
between
the
inductive
mode
and
the
subjunctive
mode.
For
example,
when
we
talk
about
wishes,
dreams
and
so
on,
we
use
the
subjunctive.
C
C
C
Okay,
so
this
table
was
created
by
a
library
called
spacey
which
which
danny
will
introduce
shortly.
Basically,
we
can
see
for
every
word
here.
Here's
the
original
sentence.
Sherlock
saw
the
man
with
the
hat.
I
changed
it
a
bit
to
avoid
ambiguity
and
for
each
such
word
we
can
see
it's
lama.
Dilemma
is
the
the
way.
The
word
appears
in
the
dictionary.
C
C
Here
you
can
see
it's
a
grammatical
label.
The
dependency
label
example
here
and
subject.
So
is
the
main
verbal
sentence
and
hence
it's
the
root
and
so
on,
and
since
many
understood
that
many
of
you
have
some
formal
background
in
computer
science
and
math,
I
suddenly
might
be
interesting
for
at
least
some
of
us
to
talk
about
dependency
as
a
as
a
graph.
C
So,
as
you
might
remember,
from
graph
series,
101
and
graph
is
basically
the
collection
of
vertices
and
edges
and
in
our
context
every
vertex
is
basically
token.
A
token
could
mean
a
word,
a
period,
a
comma,
maybe
some
sub
word
an
apostrophe
and
so
on.
Okay,
we
refer
to
it
as
talking
and
we'll
soon
talk
about
this
concept
of
tokenization.
C
C
C
E
C
C
Okay,
so
I
won't
speak
today
about
the
algorithm.
We
used
to
build
these
parse
trees,
but
let's
talk
a
bit
about
how
we
evaluate
them.
So
we
have
this
concept
of
the
tree.
Bank
freebank
basically
means
it's
a
corpus
of
sentences
and
their
respective
trees,
which
was
usually
annotated
by
humans,
and
in
order
to
evaluate
such
an
algorithm
to
build
a
tree,
we
have
two
metrics
las
and
uas,
which
stands
for
labeled
attachment
score
and
unlabeled
attachment
score,
so,
basically
for
every
pair
of
a
head
independent.
C
For
example,
danny
will
show
us
soon
spacey
and
for
each
model
space.
You
have
different
metrics
to
see
how
well
it
works
on
different
different
tasks,
and
here
I'm
sorry,
I
think
it
was
cut
off
no
here
you
can
see
it
here,
adap,
uis
and
dab
las.
This
is,
for
example,
a
model
called
trf
based
on
transformers
okay
in
english.
Basically,
you
can
see,
and,
as
you
might
have
expected,
usually
the
uas
will
be
higher
than
lis,
since
its
task
is
simpler,
doesn't
have
to
recognize
the
the
label
correctly.
C
C
Some
political
entity,
gp
geopolitical
entity,
for
example
india,
here
and
so
on.
Okay,
so
here
we
have
some
text
about
pfizer,
and
here
you
can
see
it
after
being
after
the
named
entity
recognition,
then
their
process
by
spacey.
You
can
see
here
the
different
labels,
for
example.
Pfizer
is
an
organization
biotech
an
organization.
C
C
C
C
C
C
Right
so
today
I
want
to
talk
about
very
briefly
about
two
different
methods:
different
approaches
to
tokenization
one
is
rule
based.
Basically
it
has,
if
you
see
a
comma
or
if
you
see
a
column
and
if
you
see
an
apostrophe
followed
by
us
and
so
on,
you
use
it
as
a
suffix
and
handle
such
a
special
cases.
That's
can't
vote
head
up
and
so
on.
C
Okay,
this
basically
comes
from
the
very
first
version
of
space,
and
here
you
have,
the
code
won't
go
into
the
details,
basically
check
the
prefixes
and
if
it
matches
one
of
the
cases
it
appears
and
then
checks
the
suffix
and
so
on.
The
details
are
not
very
important,
and
I
also
want
to
talk
about
a
different
approach
which
is
used
by
today
by
the
state-of-the-art
models
called
transformers
you
might
have
heard
about
about
birth.
I
think
it's
the
most
famous.
C
It's
worked
by
google
from
2019,
which
had
an
incredible
success
in
many
nlp
tasks
and
the
first
step
there
is
the
tokenization
and
use
algorithm
called
by
parent
coding,
and
basically
it
creates
it
tokenizes
it
into
sub
words
and
it's
a
statistical
algorithm
and
it
takes
that
starts
first
of
all,
where
every
single
character
is
its
own
talking
and
using
some
statistical
patterns.
These
ones
are
combined
step
by
step,
for
example,
I
might
start
with
t
h
and
e
separately,
and
then
I
start
with
the
t,
tnh
tnhe
and
so
on.
C
Okay,
so
let's
look
at
the
viper
encoding
version.
Why
do
I
have
this
crosstalk
strike
through?
I
don't
know
okay,
so
we
have,
for
example,
drug
giant
pfizer,
so
it
for
example.
Some
reason
saw
that
pfizer
is
made
of
three
different
tokens:
okay
and
the
same
for
bionda,
but
it
dealt
very
well
with
coronavirus.
C
Okay,
I'll,
let
you
think
about
it,
and
I
want
to
talk
about
another
common
step
that
we
use
in
nlp,
which
is
called
stemming.
So
if
you
remember
the
slide
about
morphology
standing,
basically
is
the
process
of
removing
fractional
morphemes
right.
So
if
we
have
a
verb
such
as
our
ranges,
we
can
normalize
it
to
a
range
and
a
very
related
concept
is
collimatization
and
when
we
remove
both
in
functional
and
relational
morphemes
example
from
arrangements
called
arrangement
and
then
arrange.
C
C
C
A
Yeah,
I
guess
that
makes
sense
if
nobody
has
any
like
pressing
questions
for
demean,
I
mean
feel
free
to
ask
them
anytime
or
put
them
in
the
chat
or
whatever
but
yeah.
Unless
there's
anything
to
say
I
guess,
maybe
it
makes
sense
to
yeah
switch
over
to
daniel.
If
you
want
to
show
us
some
examples,
closure.
B
B
B
We
need
more
layers
on
top
of
that
and
we
will
not
see
them
today,
we'll
just
play
with
the
linguistic
layer
in
a
sense,
and
the
goal
is
to
have
this
exploration,
to
get
a
sense
of
that
layer
and
also
look
into
it
as
a
beginning,
and
that
is
why
you
know
we
will
keep
asking
about
your
hopes
about
what
you
are
hoping
to
to
get
out
of
these
sessions
and
yeah.
So
that
does
this
make
sense.
Any
comments
about
that.
B
Thank
you,
yeah,
and
you
know
I
should
give
these
disclaimers.
I'm
not
an
nlp
expert,
I'm
a
statistician,
a
data
scientist
in
a
sense
I'm
a
student
of
the
mid.
I
have
been
learning
and
learning
a
lot
every
day
from
dmid
and
a
few
of
my
other
colleagues
asking
them
questions
a
lot.
What
I
will
present
here
is
playful
exploration,
enclosure
and
another
disclaimer.
B
Is
that
again
we
are
just
looking
into
basic
methods
like
in
a
sense,
a
fundamental
layer,
one
of
those
layers
that
other
algorithms
can
be
built
on
top
of
and
a
few
resources
were
looking
into
leadpython
clj.
Oh
sorry,
for
the
typo
is
a
bridge
from
closure
to
python.
We
will
use
it
today,
we're
using
we
today
and
in
other
sessions.
We
will
keep
using
python
libraries
for
this
field
of
nlp,
even
though
there
are
closure
options
and
we
will
explore
them
too.
B
It
makes
sense
in
this
field
to
have
the
best-in-class
algorithms
and
the
the
those
parts
which
are
becoming
kind
of
industry
standards
and
get
a
sense
of
them,
and
there
is
no
reason
not
to
use
them
in
closure,
and
we
know
that
closure
in
a
sense
is
a
way
to
look
into
other
runtimes
other
technologies
and
kind
of
maybe
liberate
them
and
make
them
simpler
to
use,
and
I
think
that
is
an
example
of
that.
What
what
we
will
be
seeing
today
and
tablecloth.
B
B
D
B
Of
which
is
offers
a
nice
real
world
example,
and
we
will
play
with
it,
and
I
didn't
really
thank
you
for
this.
It
really
made
sense.
I
think,
for
this
session
and
a
few
links
for
the
broader
picture.
There
is
a
nice
playlist
of
of
short
explanations
of
notions
of
nlp
for
people
of
programming
background
by
rachel
tatman.
B
It
is
part
of
this
rasa
project,
which
is
an
open
source
project
for
a
so-called
conversational
ai,
which
is
you
know,
building
chatbots
and
chat
and
such
and
they
have
actually
good
tutorials,
and
this
series
of
very
short
explanations
of
notions
and
the
practices
around
them
has
been
very
helpful
to
me,
and
data
linguist
is
something
we
should
mention.
It
is
a
closure
library.
B
Let
us
look
into
it
by
our
friend,
simon
grey.
It
is
wrapping
another
piece
of
technology
called
stanford
core
nlp.
We
should
play
with
it
in
the
future,
just
not
today,
because
for
pedagogical
reasons,
spacey
felt
so
right
and
also
because
spacey
is
so
right
for
certain
types
of
problems,
and
but
that
would
be
a
topic
for
future
sessions,
and
hopefully
simon
will
also
teach
us
about
it.
B
Of
course,
the
author
of
the
bridge,
from
closure
to
python
and
a
few
of
them,
as
well
as
ethan
miller,
has
been
exploring
this
field
of
nlp
through
interrupt
and
current
mayor
that
we
mentioned,
and
all
conversations
are
so
interesting
to
look
into
and
yeah.
So
now.
I
guess
it
would
be
good
time
to
explore,
and
maybe
what
do
you
think
kira?
Maybe
after
a
few
questions,
it
would
be
good
time
to
have
a
short
break.
A
Yeah
you
mean
before
your
demo
yeah.
That
sounds
good.
It's
been
actually
oh,
my
gosh.
It's
already
been
almost
an
hour,
so
yeah.
Why
don't
we
do
that?
If
anyone
has
any
questions,
comments
now
would
be
a
good
time.
Otherwise
yeah,
maybe
we'll
pause
for
about
five
minutes,
give
people
a
chance
to
take
a
little
break
and
come
back
around
3
35.
A
Also
available
on
the
closure,
slack
and
zulu,
there's
like
a
whole
bunch
of
places
on
the
internet,
where
all
these
people
hang
out
the
closure,
verse
they're
all
basically
like
some
form
of
closure
closureians
or
something
like
that.
If
you
google,
you'll,
be
able
to
find,
if
you
want
to
ask
like
questions
one-on-one
to
anyone
or
to
the
community
later
but
yeah,
let's,
let's
do
that.
A
B
B
You
know
spacey
in
python,
just
to
make
sense
of
it
and
yeah
I'll
show
the
screen
and
bring
the
right
window.
So
I
think
you
see
now
my
vs
code
window,
where
I
have
some
python
code.
I
could
use
emacs
but
yeah.
I
don't
know
why
I'm
here
at
ps
code,
and
so
you
see
a
few
of
the
spacey
libraries.
It
will
be
brief
just
to
have
like
a
sense
of
the
usual
way
to
use
these
things.
Spacey.
That
mentioned
is
closure.
C
Danny
yeah
in
japan,
just
just
a
small
comment
about
the
naming
scheme
of
space
models.
So
first
you
can
see
the
language
here.
It's
english
corn
means
it
includes
all
the
spacing
functionality,
for
example
the
parser
tokenizer
and
identity
condition.
So
on
what
means
it's
the
that
it
was
trained
on.
For
example,
here's
use
some
common
coral
corpus,
so
there
are
different
spacing
models
for
different
languages,
different
sizes
like
the
animation.
C
B
B
B
B
Yeah
sorry
for
this
yeah,
so
so
now
we
can
create
what
we
call
a
document
which
is
applying
our
nlp
functionality
to
a
given
piece
of
text
and
by
applying
it
we,
for
example,
get
tokenization
of
it
in
general.
We
would
get
a
certain
pipeline
or
pipeline
of
operations
applied
to
our
text
where
tokenization
is
one
of
them,
but
there
could
be
also
entity
recognition
that
is
happening
here,
but
is
not
relevant
to
our
example.
B
That
would,
for
example,
recognize
names
of
cities
right.
So
we,
this
nlp
function,
is
a
certain
pipeline
with
a
few
operations,
that's
basically
applied,
and
then
we
could
ask
our
matcher
to
match
for
this
document,
which
is
the
object
representing
the
result
of
the
pipeline
over
our
text.
And
then,
when
we
ask
for
that,
oh
yeah,
then
we
can
ask
what
the
matches
are.
Oh
sorry,.
B
B
Yeah,
thank
you.
So
our
pattern
says
we
should
have
hello
some
punctuation
and
world,
so
it
wouldn't
capture
the
second
one
right,
so
we
could
kind
of
keep
playing
and
and
and
make
this
more
flexible
and
make
this
middle
one
optional
and
such
all
of
this
is
supported.
Each
of
these
dictionaries,
as
they
call
them
in
python,
can
have
a
few
items
that
make
the
situation
precise
and,
for
example,
add
more
about
this.
So,
for
example,
we
could
say
the
lowercase
of
this
is
comma.
B
So
now
we
are
more
specific,
we're
being
more
specific
and
we're
still
capturing
this
one
example,
because
it
it
was
a
comma,
but
if
it
would
be
a
semicolon
in
our
pattern,
then
it
doesn't
capture
this
hello,
comma
world
text
right.
So
so
these
these
dictionaries
can
provide
the
details
about
what
we
expect
a
certain
token
to
be,
and
later
we
will
play
more
with
that.
So
that
is
a
little
taste
of
of
these
matches,
and
maybe
another
example.
B
B
B
B
Oh,
I
just
switched
the
the
language
model,
probably
oh
I
I
have
no
idea
what
happened
but
right.
I
expected
it
to
be
a
geographical
location.
It
was
there
earlier
today,
but
maybe
we
should
not
spend
time
on
this,
but
you
see
we
could
look
into
a
sequence
of
tokens
and
ask
about
the
properties
of
these
tokens
and
yeah
and-
and
that
is
something
we'll
look
into
more
in
closure
later,
so
that
was
a
little
taste
of
how
the
the
python
people
use
spacey.
B
B
B
This
session
is
not
about
leapbytecon
clj
and
we
will
not
dive
into
all
the
details
of
it.
Maybe
maybe
we,
if
anybody
has
questions
we
look
inside
a
little
bit,
but
we'll
just
use
it
and
enjoy
it,
and
the
idea
is
that
we
can
have
closure
functions
which
are
actually
python.
Functions
are
acting
on
python
values
and
we
will
see
that
in
a
moment,
and
so
you
see,
for
example,
you
can
require
python,
which
means
do
something
which
looks
like
the
closure
require,
but
it's
actually
requiring
python
modules
as
if
they
were
closure
namespaces.
B
We
we
have
some
namespaces
in
closure
like
spacey,
which
is
the
equivalent
of
the
python
module
called
spacing
right
and
we
will
just
use
them
and,
for
example,
we
can
take.
A
document.
Remember
document
was
what
we
get
when
we
apply
our
nlp
object
in
spacey
of
a
piece
of
text
right,
which
is
a
certain
pipeline
of
nlp
operations
acting
on
this
text
like
tokenization.
B
So
if
we
take
a
document
like
that,
then
these
documents
almost
always
sequences
of
tokens,
and
we
see
that
we
can
take
the
document
and
map
over
its
tokens
and
to
each
of
these
tokens.
We
can
see
you
know,
since
these
tokens
are
python
objects,
we
can
extract
certain
members
of
these
python
objects,
like
the
part
of
speech
that
we
saw
earlier
in
python,
and
please
don't
worry
about
it.
B
If
you
don't
understand
the
code
completely,
we
will
have
a
session
about
the
python
clj
one
day
and
we'll
dive
into
these
details,
but
I
think
you're
getting
the
sense
that
we
can,
for
example,
just
ask
for
members
of
a
python
object,
and
then
you
know
we
get
a
sequence.
Maybe
it
should
be.
Maybe
kind
of
play
with
it
a
little
bit,
so
we
can
take
a
piece
of
text
like
hello
world
and
we
can
apply
to
it
yeah.
B
B
Yeah
yeah
yeah.
Thank
you
so
much
so
I
should
have
explained
yeah.
Thank
you,
for
you
should
stop
me
more.
So
what
we
did
here
is
create
a
closure
function
called
nlp,
which
is
this
pipeline
of
spacey.
In
this
case
the
small
knowledge
base,
and
this
now
acts
as
a
closure
function.
So
we
could
say
who
could
apply
this
function
to
hello
world
and
we
can
also
write
it
this
way,
which
is
just
a
different
way
to
write
it
with
this
so-called
threading
macro
or
pipe,
as
they
sometimes
call
it
in
different
languages.
B
B
Yes,
yeah
yeah,
it
is
a
closure
function.
Well
behind
it,
we
have
a
python
function.
So
when
we
pass
a
handle
to
a
python
value
to
it,
it
would
act
on
that
python
value
and
when
we
pass
a
certain
closure
value
like
this
string,
it
would
act
in
a
sensible
way,
actually
passing
this
closure
string
to
the
python
function
and
getting
the
python
object,
which
is
the
result
and
yeah
and
it
is
so
seamless.
B
By
the
way,
all
these
notes
will
be
shared
in
a
few
days
after
some
cleanup
and-
and
I
know
I
say
we-
we-
it
is
related
to
our
exciting
efforts
to
make
the
tooling
and
set
up
more,
more
clear
and
simple
and,
and
that
is
clarifying
in
a
few
days
we'll
have
these
things
set
up
in
some
notebook-like
setup
and
yeah,
and
now
we
could
take
this
thing
and
ask
for
its
members
right.
So,
for
example,
we
can
apply
a
python,
the
famous
python
function,
that
is
called
deer.
B
B
B
C
B
Oh
so
the
tag
is
which
is
yeah
in
this
case,
maybe
not
so
much
refined
but
yeah.
It
helps
so
much
to
know
that
we
have
these
two
levels
and-
and
you
see
I
I'm
using
these
members
with
the
underscore
in
their
names.
You
also
have
those
without
the
underscore,
but
there
are
some
numerical
codes
of
these
things
so
to
make
sense
of
things
to
make
them
readability,
it's
nice
to
have
the
string,
their
presentation
and
which
is
in
some
uses
less
efficient,
but
more
readable
and
yeah.
B
B
Talk
to
data
set
this
function,
we
define
the
path,
so
it
would
look,
look
like
that
and
right.
So
it
is
a
table
right
like
they
had
in
python,
and
but
it
is
a
closure
table.
This
efficient
amazing
technical
data
set.
We
have
enclosure
which
is
wrapped
by
this
efficient,
so
nice
and
so
beautiful
tablecloth
fiber.
B
You
can
be
more
careful
about
the
way
you're
passing
values
to
python,
and
there
is
some
room
for
if
improvement,
for
efficiency
right
so
yeah.
So
now
we
have
a
function
to
match
a
little
piece
of
text
with
a
few
patterns
which
is
just
doing
whatever
we
saw
there
in
python
while
matching
so
let
us
use
that,
for
example,
I
think
I
have
yeah-
maybe
maybe
let's
copy
this
so
yeah,
maybe
like
that,
oh,
oh,
I
didn't
define
it.
B
Oh
yeah,
I
just
missed
some
function
definition,
so
you
see
where
getting
we're
getting
two
matches
for
this
text
and
this
pattern
and
we
could
remove
this
comma
and
then
we'll
only
get
one
match
right,
because
our
pattern
asks
for
some
punctuation
between
the
two
words.
The
two
words
is
it
making
sense.
B
We
are
just
wrapping
some
python
functionality,
enclosure
and
playing
with
it,
yeah,
so
yeah
and-
and
we
saw
this
displacy
python
package
that
gave
us
this
nice
dependency
plot,
and
it
also
has
a
dependency
graph,
and
it
also
has
a
certain
visual
representations,
the
representation
of
entities
that
we'll
see
in
a
moment.
So
I'm
wrapping
it
in
a
certain
way
in
closure
just
so
that
we
can
visualize
it
in
our
tools
and
yeah
or
maybe
maybe
in
a
moment
we'll
get
to
it.
B
But
first
let
us
take
some
sentence
from
our
actual
data
set
of
text
messages
about
flights
and
pass
it
to
our
nlp
python
function,
and
you
see
we're
getting
this
table
of
of
all
the
tokens
and
a
few
of
the
features
the
properties
about
them.
So,
for
example,
want
is
a
verb
which
makes
sense
right
and
any
comments
about
this.
Do
you
mean.
C
B
B
B
Yeah
and
and
timmy
again,
thank
you
for
picking
this
data
set.
It
was
so
much
the
thing
that
we
needed
for
this
session.
I
think
so
now.
Let
us
see
this
place
network
this
place
can
help
display
spacey
objects.
So,
for
example,
we
can
ask
to
display
the
dependencies
this
function
that
we
defined
here.
So
we
get
it
here
in
the
oh
yeah
maximize.
So
you
see
we
get
the
dependency
graph
of
this
sentence
with
all
the
relationship
between
words.
B
Let
us
maybe
focus
on
something
that
makes
sense
so
so
so,
for
example,
want
depends
on.
I,
because
I
is
is
the
subject
subject:
who
is
the
the
one
to
to
want
something,
and
so
on
right
and
and,
for
example,
from
depends
on
boston
and
so
on,
yeah.
So
so
we
have
this
information
and
sometimes
it
would
be
useful,
we'll
not
play
with
it
so
much
today,
maybe
maybe
a
little
in
the
exploration
we
have
later
and
another
way
to
display
a
document
is
by
coloring
the
entities.
B
So
I
think
that
is
really
nice
and
and
my
colleagues
at
work,
I
see
them
working
with
these
texts
and
they
they
coloring
that
are
coloring
everything
and
I
I
need
to
learn
from
them
to
use
this
more.
So
you
see
we
can
have
this
entity
as
a
geographical
entity
and
this
as
a
time
entity
and
so
on,
and
this
update
right
great.
B
So
now
let
us
go
about
the
problem
that
we
are
looking
into.
We
will
not
have
any
great
solution
to
our
problem,
we'll
just
explore
it
a
little
bit
so
I'll
have
this
other
closure
namespace
and
in
this
namespace
we
will
look
into
this
artist
dataset
that
we
looked
into
in
the
beginning-
and
maybe
maybe
let
us
have
a
reminder
of
what
it
is
so
right.
B
Or
so
on,
and
it
is
given
in
different
formats,
and
one
of
them
is
the
so-called
rasa
nlu
json
format,
which
is
a
way
to
represent
a
set
of
texts
and
some
information
that
needs
to
be
ex
extracted
from
them.
So
we
see
that
these
texts
were
annotated
with
some
information
that
can
be
extracted
from
them.
This
was
probably
done
by
a
human
saying.
B
The
intent
of
this
text
is
looking
for
a
flight
and
intense
this.
You
know
general
idea
of
looking
into
pieces
of
text
and
somehow
simplifying
their
meaning
very
much
in
a
gen
in
a
very
generalized
way,
like
categorizing
them
into
a
few
different
meanings
general
meanings
that
people
could
be
talking
about
not
very
details
but
very
much
simplified,
and
it
is
a
very
common
practice
in
nlp
systems
to
build
a
system
by
first
recognizing
some
intents
and
then
going
more
detailed
about
it,
and
you
know
that
is
some
of
the
practice.
B
I'm
learning
these
days
from
dmid
and
my
colleagues
and
you
see
this
was
annotated,
probably
by
humans
as
a
way
to
test
the
ability
of
a
certain
algorithm
to
recognize
those
intents.
So
these
data
would
typically
be
used
for
training,
machine
learning,
algorithms
or
some
other
algorithms
and
testing
them.
So
we
have
a
training
and
a
training
set
of
data
like
that
and
a
testing
set
of
data
like
that
and
this
practice
of
training
and
testing
so-called
supervised
learning.
B
We
will
have
more
of
it
in
future
sessions
and
today,
we'll
just
look
into
the
training
set
and
play
with
the
ability
to
maybe
maybe
recognize
some
of
these
details,
and
you
see
we
don't
only
have
intense
but
actually
detailed
entities.
So,
for
example,
in
this
text
we
have
this
entity
of
the
city
name
they
wish
to
be
flying
from,
and
this
is
charlotte
right,
so
yeah
and
you
also
have
the
so-called
span
of
where
it
exists
in
the
text.
B
And
so
these
are
the
data,
and
maybe
I
should
I'm
kind
of
stating
the
obvious
thing
that
when
we
see
something
like
that
in
closure,
as
as
closure
developers,
we
have
this
happiness
about
this
moment,
because
we
know
that
working
with
jason
in
closure
is
such
a
joy
and-
and
I
think
it
also
kind
of
applies
to
spacey
and
the
whole
ecosystem.
Around
spacey,
because
it
is
also
very
much
organized
around
data
structures,
nested
data
structures
like
these
patterns,
which
is
something
that
we
like
and
enjoy
to
create
and
manipulate
enclosure.
B
So
you
see
we
have
the
intent
in
one
column,
flight
and
the
text,
and
then
the
nice
thing
is
that
we
can
have
all
these
entities
in
a
in
a
kind
of
in
a
data
set
inside
this
column
and
in
in
table
cloth
and
the
underlying
the
camel
data
set
library.
This
is
very
much
efficient.
This
ability
to
have
data
sets
inside
datasets
and
so
on
and
and
yeah
and
and
we
have
these
data,
does
it
make
sense
that
these
texts
were
annotated
with
intent
and
entity
information?
Maybe
we
should.
B
Reorder
the
columns
right
and
never
mind
yeah,
you
see
yeah,
so
that
is
the
text
and
it
has
intent
and
entities,
and
that
is
the
information
we
are
hoping
our
algorithms
to
learn
to
recognize
right
and
yeah.
So
so
now
we
have
both
the
training
set
and
and
the
test
set,
which
is
this
common
practice
of
dividing
our
data
into
something
we're
practicing
and
learning
from
and
the
part
where
we
only
test
our
algorithms
without
even
looking
inside
ever
right
and
and
typically
in
future
sessions.
B
D
B
Take
the
first
text
and
write
this
text
and
turn
it
into
an
nlp
document
using
our
nlp
function
and
then
into
a
nlp
data
set
right,
so
we're
just
practicing
what
we
did
earlier
just
with
this
sentence
again
and
we
see
it
so
this
is
not
new
and
and
now
maybe
let
us
visualize
it.
So
let
us
go
to
our
browser
with
this
clay
tool
that
will
discuss
more
in
future
sessions.
B
Probably
so
you
see,
it
is
very
useful
to
have
these
entities
colored,
because
these
entities
are
now
recognized
by
spacey
right.
So
so
we
have
here
two
kinds
of
entities
in
in
the
original
data
set.
We
had
those
entities
that
were
annotation
annotated
over
the
data
by
humans
as
the
target
for
algorithms
to
learn.
So,
for
example,
we
have
the
from
and
to
city
and
the
departure
and
arrival
time
right,
but
spacey
also
has
the
ability
to
recognize
entities,
but
these
would
be
a
bit
different,
spacey
wouldn't
know
about
from
and
to
cities.
B
B
C
Just
one
comment
about
the
data
set:
you
might
be
wondering
why
the
names
of
the
cities
are
not
capitalized
and
why
there
is
no
punctuation,
and
the
reason
is
that
it
comes
from
speech
recognition.
So
basically
users
called
some
number
and
then
it
was
transcribed
automatically
and
that's
why
we
don't
have
this
information,
which
makes
our
nlp
task
a
bit
more
challenging.
B
Oh,
thank
you
so
much
for
this.
I
I
forgot
this
detail
completely
and
yeah
yeah
fantastic.
Thank
you
yeah.
So
you
see
we
could
visualize
it
and
and
and
have
this
look
into
entities
and
we
can
also
visualize
dependencies
right
so,
for
example,
this
yeah
we
already
looked
into
this,
and
but
here
you
see
how
fun
it
is
to
use
this
hiccup
notation
of
closure
that
allows
you
to
create
html.
B
You
know
in
this
joyful
way
with
list
items
and
such
and
just
organize
your
visual
information
yeah
easily,
and
it
just
works
with
these
python
libraries
as
long
as
you
just
wrap
them
correctly
and
yeah
and
and
yeah.
Maybe
let
us
look
into
the
different
different
intents
we
have
so
right.
We
have
mostly
the
flight
intent,
but
also
others
and
yeah,
sorry
and
yeah,
and
now
yeah
I'll
skip
a
little
bit.
I
think,
and
now
I'll
actually
add
this
departure
time
column.
B
B
You
see
in
some
texts
the
human
annotation
said
yeah.
This
text
did
have
some
departure
time
in
it
right.
This
one
did
have
this
8
38
a.m,
all
right.
So
oh,
what
happened?
Yeah
yeah,
so
I
just
extracted
that
as
a
separate
column,
so
that
it
would
be
easier
to
reason
about
it
and
doing
that
with
tablecloth
is
so
fun.
That
is
not
the
topic
of
today,
but
it
was
just
so
joyful
to
have
these
internal
data
sets
and
extracts
and
values
from
them
and
yeah.
Anyway.
B
B
How
often
does
it
happen
that
it
is
not
nil
or
that
it
is
new
right,
so
it
is
nil,
most
of
the
time,
but
oh
355
times
it
was
not
nil,
so
it
actually
happens
in
the
training
set
that
sometimes
the
annotations
do
contain
this
departure
time.
So
we
could
hope
to
have
enough
information
to
learn
something
about
how
to
recognize
departure
times
in
these
texts,
and
you
know
I'm
just
trying
to
give
a
taste
of
how
it
is
to
play
with
tablecloth
and
the
bridge
to
python
and
such,
and
it
is
okay.
B
If
you
don't
understand
all
these
examples,
these
will
be
topics
of
future
sessions
and
and
I'm
happy
to
discuss
it
later
with
anybody
who
is
curious
and
yeah.
Maybe
now
let
us
look
into
a
few
examples
where
we
have
departure
tax.
So
you
see
I'm
taking
the
training
step
after
this
preprocessing
and
I'm
selecting
the
rows
where
the
departure
time
is
not
nil,
which
means
those
rows
where
the
humans
annotating.
The
text
said
yeah.
B
Of
course
we
have
a
problem,
because
we
need
this
to
discover
this
as
a
departure
time,
but
in
the
data,
when
we
apply
spacey
spacey
would
maybe
realize
it
is
a
time
value,
but
not
a
departure
time,
because
it
doesn't
know
what
it
is
and
our
goal
is
to
to
define
some
algorithm
to
do
that
and
we'll
just
play
with
it
a
little
bit
of
course,
and
does
it
make
sense?
I'm
you
know
setting
up
the
problem
and
we'll
go
about
these
things
more
and
more
with
different
methods
in
the
future.
B
B
We
are
looking
for
the
words
before
after
or
at
and
then
for
a
time
entity
right.
So
let
us
see
if
this
pattern
will
match
this
text
and
sorry
for
the
nesting
here
we
are
kind
of
switching
between
the
tablecloth
way
and
the
plane
closure
way.
So
this
code
might
look
confusing
a
little
bit,
and
maybe
this
way
it
is
a
bit
more
readable,
but
maybe
maybe
I
hope
you
don't
care
about
it.
B
The
second
token
is
a
time
token,
like
6pm
the
first
token,
or
I
think
six
in
this
case,
because
we
recognize
it
yeah,
never
mind
about
this
technical
detail,
but
because
in
space
you
could
approach
it.
You
could
unify,
unify
these
two
as
one
entity
talk
one
token
of
time
entity,
but
you
could
also,
depending
on
your
pipeline.
B
You
could
also
define
them
as
two
separate
token
or
time
entity,
and
it
depends
on
your
choice,
how
you
would
represent
it,
and
I
think
now
we
are
representing
them
as
two,
if
I
remember
correctly,
or
maybe
not
yeah,
but
never
mind,
and
what
we
are
asking
for
in
the
previous
token
is
that
it
would
be
some
token
where
the
lemma
of
this
token
well
lemma,
is
this
simplification
of
the
word
it
into
some
common
sense.
That
is
kind
of
avoiding
different
variations
of
certain
words.
B
The
lemma
would
be
one
of
these,
and
we
see
that
this
pattern
does
find
it
in
our
text
and
in
this
case
we
will
write
if
we
wanted
to
use
this
button.
With
this
text,
we
will
write
about
recognizes
recognizing
these
two
entities
as
something
about
departure
time,
but
of
course
it
could
be
that
in
some
other
text
we
would
not
be
right
and
they
would
be
actually
the
arrival
times.
So
we
need
a
better
pattern.
E
B
B
D
B
World
and
we
want
to
make
the
comma
optional,
so
I
think
it
is
up
as
the
mid
set-
and
I
think
star
would
make
sense
here
or
maybe
question
mark,
because
we
want
to
allow
it
once
so.
This
does
work
because
we
made
the
comma
optional
and
question
mark
is
of
the
same
meaning.
You
would
see
in
regular
expressions
for
those
who
care
about
that
right.
It
is
something
that
says
you
could
have
that
thing,
maybe
once
maybe
none
right,
and
so
this
works,
but
without
making
it
optional,
it
would
not
work.
B
This
tool
is
useful
but
limited.
You
will
see
it
in
small
projects
as
a
way
of
defining
patterns
and
recognizing
things
in
this
very,
very
detailed
way,
and
that
would
be
very
much
fun
to
do
enclosure.
I
think
you're
getting
a
sense
of
it
because
of
these
tiny
little
nice
data
structures,
but
it
would
also
be
useful
as
a
first
layer
in
a
sophisticated
natural
language
processing
pipeline.
That
has
many
other
things
throughout
the
pipeline.
B
B
A
Yeah
yeah
thanks
so
much
for
going
through
that
all
daniel.
It's
super
interesting
yeah
anyway.
I
think
that
sounds
good
to
me.
We
have
a
couple
minutes
left
until
the
official
time
we
said
we
would
wrap
up.
So
if
anyone
has
any
questions,
feel
free
and
also
like
daniel
mentioned,
we
can
stick
around
for
a
little
bit
longer
if
you
would
like
you're
welcome
to
as
well.
F
D
F
F
Experience-
and
that
is
really
it
comes
from
this
fact
that,
especially
in
the
area
of
data
science
yeah
as
a
data
scientist,
what
you
want
to
do
you
want,
you
have
a
problem.
You
want
to
read
the
paper
and
you
want
to
do
what
is
written
in
the
paper
and
if
the
paper
says
use
python
algorithm
x,
then
you
don't
want
to
reprogram
something
else.
You
want
to
use
that
algorithm.
F
So,
basically
you
are
not.
You
don't
have
a
choice
in
most
of
the
cases
to
say:
okay,
I
use
something
different
similar,
but
if
it
exists
in
jo
in
in
java
or
closure,
because
then
you
change
the
algorithm
basically
right-
and
that
is
my
argument.
Why
is
this
absolute
important
to
stay
in
the
in
the
interop?
A
Yeah
yeah
for
sure,
and
just
in
terms
of
the
amount
of
effort,
I
think
you
know
it
just
doesn't
make
sense
to
be
re-implementing
super
niche
things
in
a
whole
bunch
of
different
languages,
like
you
said
so,
and
yeah
python
is
clearly.
You
know
the
kind
of
the
dominant
language
that
a
lot
of
these
machine
learning
things
have
been
developed
and
so
yeah,
it's
cool
to
be
able
to
see
how
to
use
them.
F
F
Already
in
our
areas,
for
example,
it
is
feasible
to
to
redo
something
like
pandas.
If
you
talk
about
data
data
manipulation,
because
table
close
in
my
view
is
very
close
to
that.
No,
but
if
you
want
to
do
the
same,
the
same
for
the
machine
learning
algorithms,
that
is
impossible
because
there's
such
a
huge
amount
of
things
and
nobody
wants
to
re-implement
them.
That
is
why
yeah
there's,
of
course,
a
little
bit
of
of
duplication
between
r
and
python,
because
they
are
both
very
big
ecosystems.
D
F
F
F
F
You
cannot,
I
think,
it's
impossible
to
think
about
something
which
wraps
that
completely,
because
it's
changing
so
quick.
You
cannot
change
your
wrapper
as
quickly
as
it's
changing,
but
it
is
true
that
lib
python
clj
is
very
nice
to
use
itself,
so
it
is
really
fun
to
use
it
and
do
these
little
these
little
wrapping
functions
as
part
of
your
work,
but
only
the
piece
you
want
to
use
at
a
given
time,
but
how
you've
done
it
today.
So
I
think
this
is
the
right.
The
right
way
to
go.
B
Yeah
and
at
the
same
time,
psychology
ml
is
offering
wrappers
of
a
few
of
those
amazing
collections
like
scikit-learn
and
this
smile
java
library
and
this
deep
java
library
right.
Maybe
these
are
the
three
main
collections,
and
so
smart
psychology
has
a
lot
of
effort
of
actually
writing
raptors
right.
F
F
So
basically
I
just
needed
to
do
a
wrapper
for
the
interface,
not
for
the
algorithms,
so
that
is
feasible
and
it's
similar
to
smile.
There
was
a
little
bit
of
a
problems
with
smile
in
the
beginning
that
not
all
smile
machine
learning
models
had
the
same
interface.
So
there
was
a
little
bit
one
of
the
reasons
why
there
was
a
little
bit
of
work
there,
but
that
is
not
the
case,
so
that
is
feasible.
F
If
you
have
a
big
library
in
all
the
the
machine
learning
or
a
lot
of
algorithms
have
the
same
interface,
it
is
feasible
to
have
a
wrapper
for
the
interface,
so
you
get
all
the
algorithms
at
the
same
time,
but
this
is
something
which
exists
for
machine
learning
as
such,
because
the
table
is
the
input
so
so
table
are
data
so
because
they
all
take
a
table
as
input
and
they
give
the
model
out.
But
I
don't
see
this
for
the
nlp
part,
because
all
nap
models
or
approaches
have
different,
very
different
inputs
and
outputs.
F
So
you
need
to
do
that
needed
to
do
the
wrapping
one
by
one,
and
that
is,
I
think,
not
feasible
in
my
view,
but
that
that
is.
That
is
how
I
see
the
difference
so
in
in
in
smile
it
is
or
an
eska
learn.
It
was
indeed
possible
because
escalante
made
the
effort
that
all
algorithms
have
the
same
python
interface,
so
the
wrapper
was
simple
and
it's
stable,
it's
stable
as
well.
F
B
Thank
you,
it
helped
so
much
yeah,
maybe
kira.
Maybe
should
we
in
a
moment
stop
the
recording
and
say
goodbye
to
the
recording?
Maybe
you
or
dimit
have
some
conclusion
to
say.
I
only
want
to
say
thank
you.
It
was
you
know
this.
This
group
is
so
supportive
of
these
explorations
and
such
a
joy
to
have
these
discussions.
A
Yeah
yeah,
this
was
a
cool
one
for
sure,
so
yeah
I
don't
think
I
must
add
just
yeah
big
thanks
to
demean
for
joining
us
and
teaching
us
a
little
bit
about
nlp
and
thanks
to
daniel
for
the
demo
and
yeah.
I
guess,
like
I
mentioned
myself
and
daniel,
and
I'm
not
sure
too
sure
about
to
me,
but
probably
we're
all
available
like
around
the
internet.
If
you
have
any
questions
or
like,
if
you
have
interest
in
this
kind
of
meetup,
we
could
potentially
make
this
a
recurring
event.
A
I
think
at
the
moment
we
don't
have
plans
for
like
the
next
one
anytime
soon,
but
there
are
a
lot
of
study
groups
that
meet
like
every
two
weeks
or
every
month
or
something
like
that,
and
I
think
this
would
be
a
really
interesting
topic
for
a
lot
of
people.
So
and
definitely
this
like
closure,
python,
interop
stuff
will
be
featured
in
the
the
closure
cookbook
and
the
data
science
course
that
is
kind
of
sort
of
coming
up.