►
Description
In this 15-minute all audiences-focused demo, James Eccles shows the 4-step UI based process for building a Syntax Document Understanding Model from scratch. Steps include uploading example files, classifying them, defining entity extractors, and publishing the model to libraries. Train your model by identifying (labeling) correct and not correct attributes and then use text explanations (hints) to help Syntex look for specific terminology in document. Run documents through the model and see metadata extracted from files in a new Document View in your document library. This PnP Community demo is taken from the SharePoint monthly community call recorded on May 11, 2021.
Demo Presenter - James Eccles (Microsoft) | @jimdeccles
Supporting documents:
• Web - SharePoint Syntex | aka.ms/SharePointSyntex
• Blog – Syntex blog | aka.ms/SharePointSyntex/blog
• (Soon) Discussion – Yammer Syntex Community | aka.ms/SharePointSyntex/community
• Roadmap – Syntex Roadmap | aka.ms/SharePointSyntex/roadmap
• Documentation - Introduction to Microsoft SharePoint Syntex | aka.ms/SharePointSyntex/documentation
• Documentation – Scenario: Manage contracts using a Microsoft 365 solution | https://docs.microsoft.com/en-us/microsoft-365/contentunderstanding/solution-manage-contracts-in-microsoft-365
• SDK - Working with SharePoint Syntex | https://pnp.github.io/pnpcore/using-the-sdk/syntex-intro.html
• (Soon) Samples - GitHub: https://github.com/pnp/syntex-samples
• Learning – Virtual Hub | https://adoption.microsoft.com/virtual-hub/search?_sft_vh_product=sharepoint-syntex
Learn more:
• Learn more about the Microsoft 365 Patterns and Practices community at: http://aka.ms/m365pnp
• Visit the Microsoft 365 Patterns and Practices community YouTube channel: https://aka.ms/m365pnp-videos
A
What
I'm
going
to
show
now
is
how
to
build,
publish
and
publish
a
document
understanding
model
document.
Understanding
models
are
really
suited
towards
unstructured
data.
So
in
the
example
I'm
going
to
share
we'll
look
at
benefit
change
notices.
Here's
an
example
of
the
type
of
file
that
we're
going
to
build
a
model
to
understand.
You'll
see
it's
it's
unstructured,
I
can't
say
for
the
data.
I
want
to
pull
as
metadata
where
that
might
appear
in
the
document.
A
I'm
going
to
begin
in
the
content
center,
which
is
the
new
site
template
that
sean
was
just
describing
and
use
the
content
center
to
build
my
document
understanding
model
and
then
publish
it
to
one
or
more
sharepoint
libraries,
I'll
I'll
talk
about
the
ui
way
of
publishing
models
in
in
my
section,
the
demo
and
then
we'll
talk
about
some
of
the
options
over
apis
and
through
powershell
with
burp.
In
a
couple
of
minutes,
so
let
me
start
off
by
creating
a
new
document.
Understanding
model.
A
A
We've
also
got
the
option
here
to
associate
a
retention
label
with
my
model,
so
in
this
case
I'll
grab
the
confidential
label.
So
whenever
a
benefit
change
notice
is
processed
by
syntax.
Using
this
model,
it
not
only
will
classify
it
and
extract
the
metadata.
It
will
stamp
the
confidential
label
on
there
as
well
to
preserve
the
life
cycle
of
the
file.
A
When
we're
building
a
document
understanding
model,
there
are
four
steps
that
we
walk
through.
So
first
we
upload
a
training
set
of
files.
Then
we
build
something
called
our
classifier
and
that's
going
to
say
when
a
file
is
executed
against
this
model,
is
it
a
benefit
change?
Yes
or
no,
or
is
it
another
file
type,
such
as
a
contract
or
a
statement
of
work?
How
can
we
make
the
determination
whether
it
is
or
is
not,
that
file
type?
A
But
again
with
machine
learning,
where
you
might
be
looking
at
hundreds
or
thousands
of
labeled
files
when
we're
working
with,
because
I've
got
a
file
open,
that'll,
be
why
we're
working
with
sharepoint
syntax.
We
only
need
a
small
number
of
files,
a
minimum
of
five
and
what
we're
going
to
do
through
the
process
is
less
around
getting
ai
to
analyze
the
data
set
and
find
patterns.
We're
going
to
use
this
small
data
set
to
walk
through
an
interactive
training
process.
So
the
idea
is
that
I
is
an
expert
in
this
type
of
content.
A
I'm
going
to
use
the
model
to
impart
my
knowledge,
I'm
going
to
teach
the
ai
and
the
ai's
then
going
to
be
able
to
understand
the
content
going
forwards
as
a
bare
minimum
minimum.
We
need
five
files
to
teach
a
document,
understanding
model
as
a
mixture
of
positive
examples
and
negative
examples.
In
my
case,
I've
got
14
files,
so
12
examples
of
a
benefit,
change
and
two
statements
of
work
to
give
the
ai
an
example
of
something
that
isn't
a
benefit
change.
A
Now
I've
got
my
data
set
I'll
move
on
to
the
classifier
and
when
we
work
with
document
understanding
models,
I
really
like
to
use
the
language
of
teaching.
So
imagine,
you're,
teaching
a
human
being
and
the
same
steps
you'll
go
through
what
we
do
when
we're
teaching
the
ai
when
we're
teaching
syntax
how
to
build
a
model.
A
So
yeah
I've
got
pretty
small
children.
So
when
I'm
teaching
them
how
to
do
things
like
recognize
shapes
or
animals,
first
I'll
show
them
examples
and
then
you'll
explain
the
difference
between
them.
So
maybe
maybe
I'm
teaching
the
difference
between
a
dog
and
a
cat.
My
children
are
very
small,
so
that's
the
level
we're
at
I'll
show
them
some
pictures
of
dogs,
some
pictures
of
cats,
so
they
can
identify
the
difference
and
then
maybe
I'll
I'll
start
to
explain
the
difference.
So
dogs
might
tend
to
be
bigger
or
cats
can
climb.
A
Trees
and
dogs
can't
I'll
I'll
point
out
the
attributes
of
both
to
help
them
work
out
for
themselves
in
the
future.
We're
going
to
do
exactly
the
same
with
syntax.
First,
we're
going
to
give
examples
in
this
case
we're
going
to
label
some
of
the
documents
and
then
we're
going
to
explain
how
we
came
to
that
conclusion.
We'll
use
what
we
call
explanations
to
provide
additional
context
and
syntax
will
use
a
combination
of
those
explanations
and
the
labeling
that
I
do
to
build
the
model.
A
I'm
looking
at
the
labeling
interface
at
the
moment.
On
the
left
hand,
side
you'll
see
the
files
that
I
have
in
my
training
set
and,
on
the
right
hand,
side
a
preview
pane
of
the
file
now
for
the
preview
pane,
we've
hidden
all
of
the
rich
text,
all
the
rich
formatting
from
the
document
we're
only
interested
in
the
text
when
we're
training,
syntax
so
to
make
the
training
process
easier.
We've
we've
hidden
out
all
that
rich
detail.
A
I
could
hit
the
view
original
button
to
get
into
that
original
version
of
the
file
at
any
time,
but
for
now
we're
only
interested
in
the
text
and
it's
a
pretty
simple
step
for
labeling.
I
need
to
go
through
at
least
five
of
these
files
and
say
yes
or
no,
whether
or
not
it
is
an
example
of
a
benefit.
A
A
Next
is
training
so
again
I've.
Given
my
ex
my
examples
now,
I
need
to
show
my
thinking
show
my
working
out
and
the
explanations
are
interesting
when
I
explain
this
to
customers
oftentimes,
they
get
bogged
down
with
other
technologies,
they're
used
to
and
trying
to
apply
some
of
those
thinking
patterns
to
syntax,
so
they'll,
think
of
explanations
as
being
regular
expressions
or
different
types
of
query
statements
you're
having
to
build
rules
that
will
say
exactly
where
text
appears
within
a
document
and
the
reality
of
syntax
is
that's.
That's
not
the
case.
A
What
explanations
are
are
hints
they're
suggestions
of
how
the
ai
might
identify
documents
or
the
attributes
in
them
again
think
about
how
you're
teaching
a
person
you
don't
necessarily
explain
to
a
person
how
many
characters
into
a
document.
A
particular
phrase
is
you'll
tell
them
where
you
might
identify
a
phrase
or
the
type
of
language
you'll
see
in
a
document
that
tells
you
that
it's
a
benefit,
change
or
a
contract,
and
so
on.
A
A
A
On
the
bottom,
you
can
see
for
each
of
the
files,
I've
labeled
we've
got
the
predicted
label
and
an
evaluation.
So
the
evaluation
is
what
the
ai
would
do
based
on
the
instructions
it's
been
given.
So
does
it
agree
with
me
yes
or
no?
As
the
trainer
and
you'll
see,
we
have
five
out
of
five
matches
so
for
now
my
model
is
100,
accurate.
So
for
the
classifier,
the
ai
agrees
with
what
I
would
do
as
a
human
expert.
A
I
can
move
to
the
testing
stage
and
use
the
rest
of
my
data
set
to
do
some
more
testing.
That
was
a
pretty
simple
example:
let's
move
to
an
extractor
to
show
you
something
a
little
bit
more
complex
like
ramp
up
the
complexity
just
a
little
bit
when
I
create
extractors
they
map
to
the
columns
which
will
exist
on
the
sharepoint
library
that
I
publish
to
so
I
can
create
many
extractors.
Each
one
will
be
a
separate
column
in
this
demo.
I'll
just
create
one
for
time,
but
you'll
see
some
of
the
ideas
here.
A
So
again
I
can
map
this
to
either
a
new
or
an
existing
column.
I
can
set
what
data
type
I'm
going
to
be
extracting
from
the
particular
file
in
this
case,
let's
grab
the
effective
date
of
the
contract
change,
and
this
might
be
useful
once
we
extract
the
metadata
for
doing
things
like
triggering
some
automation,
so
we
might
need
to
do
notifications
based
on
that
date,
we
might
want
to
sort
or
filter
or
build
a
view
formatter
on
the
library
which
is
using
this
date
column
to
provide
some
improved
navigation
and
readability.
A
And
move
on
to
my
training
stage
before
I
start
to
build
some
explanations:
just
notice.
On
the
right
hand,
side
we've
still
got
a
blue
highlight
on
the
on
the
preview,
and
that
shows
me
what
I
labeled
in
the
first
step.
As
I
start
to
layer
on
explanations,
we'll
also
see
a
green
highlight
on
the
on
the
preview
and
the
green
highlight
shows
what
the
ai
would
extract
from
this
document,
and
we
want
to
see
the
two
things
match
up.
A
A
You'll
see
that's
evaluating
based
on
the
instructions
I've
given
at
the
bottom
left,
and
what
you
might
be
surprised
to
see
now
is
that
we
have
four
out
of
five
mismatches.
So,
given
that
I've
added
a
date,
you
might
be
pretty
surprised
by
that,
because
there
is
a
date
in
this
file.
I've
told
syntax
that
it's
looking
for
a
date,
so
why
hasn't?
It
picked
out
this
particular
date,
and
this
reinforces,
hopefully
the
the
idea
that
we're
not
building
regular
expressions
or
query
statements.
A
These
are
in
the
same
way
that
you
teach
a
person
hints
at
how
to
find
the
right
data.
In
this
case,
I've
told
syntax
that
it's
a
date
it's
looking
for,
but
there's
no
context
here
to
say
which
date
in
this
file
am
I
looking
for.
How
do
I
know
this
is
the
effective
date
rather
than
the
date.
The
document
was
written.
A
A
We've
got
the
green
and
the
blue
highlights
matching.
If
I
look
at
my
mismatch,
though,
what
we
can
see
is
that
it
thinks
there's
an
effective
date
in
the
statement
of
work,
so
we
need
to
give
additional
cues
additional
linguistic
clues
and
hints
to
the
ai
in
order
for
it
to
disambiguate.
So
again
we
can
add
more
explanations,
either
using
the
automatic
phrase,
lists
or
manual
phrase
lists.
A
We
now
have
accuracy
across
all
of
my
five
test:
files
expand
out,
look
at
the
rest
of
the
training
set
and
then
finally,
we
can
publish
this
to
a
sharepoint
library.
Now
I'm
going
to
cheat
for
time
and
go
to
an
example
that
I
created
earlier
on.
So
this
is
using
the
same
test
set,
but
I've
created
a
number
of
more
a
number
of
more
extractors
here
as
well.
B
A
We've
also
created
a
new
view
for
this
document
library,
so
we've
changed
the
default
view
to
be
one
based
on
benefit
changes.
The
model
I've
created,
which
shows
all
those
columns,
as
well
as
the
content
type
and
the
confidence
score
and
classification
date
time,
which
are
additional
system
columns.
We
generate
within
sharepoint
syntax
now
over
the
next
couple
of
minutes.
A
These
files
will
process
and
we'll
see
the
the
results
from
that
generally
takes
between
two
and
ten
minutes
for
content
to
properly
process
through
through
sharepoint
syntax
I'll
show
you
an
example
of
one
I
made
earlier,
which
shows
all
of
the
metadata.
That's
been
automatically
extracted
from
each
of
those
files,
the
confidence
the
syntax
hat
syntax
had
in
extracting
that
and
the
timestamp
of
when
the
extraction
ran.