How to Split Re ursive Automata

Isabelle Tellier

Description

Niveau: Secondaire, Lycée, Terminale
How to Split Re ursive Automata Isabelle Tellier LIFL - Inria Lille Nord Europe university of Lille isabelle.tellieruniv-lille3.fr ?? In this paper, we interpret in terms of operations applying on extended nite state automata some algorithms that have been spe ied on ategorial grammars to learn sub lasses of ontext-free languages. The algorithms onsidered imple- ment spe ialization strategies. This new perspe tive also helps to understand how it is possible to ontrol the ombinatorial explosion that spe ialization te hniques have to fa e, thanks to a typing approa h. 1 Introdu tion There are often several ways to represent a language: it is well known that every regular language an be spe ied either by a regular grammar or by a deter- ministi nite state automaton. Context-free languages an also be spe ied by dierent kinds of devi es. In re ent previous papers [17, 18?, we have shown that some lasses of ategorial grammars (CGs in the following), generating ontext- free languages, ould easily be represented by a family of extended automata alled re ursive automata (RA). This translation allowed to exhibit onnexions between two previously distin t approa hes of grammati al inferen e from posi- tive examples: the one used in [3, 13, 14? to learn CGs, and the one used to learn regular grammars represented by nite state automata [1, 10?.

• been briey


• introdu ing


• nal vo


• transition labelled


• free ontext-free


• ba rules


• state automata


• q? ?


• nite state


• spe ied either

Sujets

• Première
• Tellier

Informations

• Publié par: profil-zyan-2012
• Nombre de lectures: 11
• Langue: English

Extrait

⋆⋆
⋆⋆

e
Automata
Isab
ed
w
In
elle
learn
T
wn
ellier
re-
LIFL
ecial-
-
in
Inria
.
Lille
pro
Nord
set
Europ
e
e
understand
univ
oth
ersit
A
y
h
of
emplo
Lille
in

e
fr
gies
Split
othesis
to

In
it
this
CGs
pap
generalization
er,
a
w
een
e
CGs
in
nite
w
that
orted
this
b
terpret
in
regular
terms
automata
of
b
op
alization
erations
generalization
applying
texts
on
similar.
extended
to
nite
e
state
os-
automata
the
some
general
algorithms

that
onstr
ha
h
v
the
e
that
b
whic
een
understand
sp
help
ecied
sp
on
although

ha
grammars
enden
to
learn
learn
hand
sub
resen

the
of
e

een
text-free
to
languages.
e
The
generalit
algorithms
w

ANR
imple-
represen
men
nite
t
℄
sp
p
e
b

a
str
ate
ate
particular,
gies
erators
.
oth
This
ere
new
b
p
w,
ersp
an
ectiv
the
e
for
also
str
helps
om
to
examples
understand
h
ho
h
w
to
it
grammar
is
grammars)
p
example
ossible
a
to
whic

the
trol
un
the


grammar.
binatorial
w
explosion
translation
that
to
sp
has
ecialization
b

family
hniques

ha
b
v
family
e
strategies.
to
of

used,
thanks

to
e
a
osed
t
in
yping
kgrounds:


h.
the
1
and
In
grammars
tro
b

automata
There
hand
are
mo
often
w
sev
has
eral
prop
w
but
a
CGs.
ys
er,
to
the
represen
its
t
.
a
w
language:
partly
it
y
is
CroT
w
grammars
ell
ted
kno
y
wn
state
that
[1,
ev
This
ery
as
regular
ossible
language
ecause

oth
b
y
e
gener
sp
str
ecied
gy
either
In
b
the
y
op
a
used
regular
b
grammar

or
w
b
sho
y
to
a
e
deter-
No
ministic
w
nite
w
state
t
automaton.
apply
Con
same
text-free

languages
sp


also
ate
b
fr
e
p
sp
itive
ecied
.
b

y
strategies,
dieren
initial
t
yp
kinds
is
of
o

a
In
(or

of
t
and
previous
h
pap
is
ers
as
[17,

℄
aint
w
h
e

ha

v
space,
e
til
sho
is
wn
to
that
target
some
W

sho
of
here

the
grammars
of
(CGs
in
in
RA,
the
h
follo
help
wing),
to
generating
etter

the
text-
of
free
strategies,
languages,
also

to
easily
etter
b
the
e
of
represen
ecialization
ted
As
b
matter
y

a
barely
family
sp
of
ization
extended
hes
automata
v

b

prop
e
indep
automata
tly
(RA).
b
This

translation
to
allo
sub
w
of
ed
in
to
one
exhibit
[16],

to
b
regular
et
rep-
w
ted
een
y
t
state
w
in
o
other
previously
[11].

rst

v
hes
to
of
ards


inference
b
from
briey
p
osed
osi-
[19],
tiv
limited
e
unidirectional
examples:
In
the
pap
one
w
used
generalize
in

[3,
to
13,
full
℄
y
Ho
to
learn
This
CGs,
ork
and
as
the
supp
one
b
used
the
to
MDCO
learn
alB
S ∈B Cat(B) B⊂Cat(B)
A,B ∈ Cat(B) A/B ∈ Cat(B) B\A∈ Cat(B)
/ \
Σ B S
G Σ×Cat(B) hv,Ci∈G
C ∈ Cat(B) v ∈Σ
∀A,B ∈Cat(B)
FA A/B B →A
BA B B\A→A
FA BA
G
+L(G) {w = v ...v ∈ Σ |∀i∈{1,...,n} ∃C ∈ Cat(B) hv ,C i∈1 n i i i
∗G C ... C → S}1 n
∗→ →
FA BA w ∈ L(G)
G FA BA
B ={S,T,CN} T
CN Σ ={John,runs,a,man,fast} G =
{hJohn,Ti,hruns,T\Si,hman,CNi,ha,(S/(T\S))/CNi,hfast,((T\S)\(T\S))i}
from
CGs
d
the
,
rigid
prop
learn
of
and
e
for
to
any
b
to
b
wing
2
allo
new
[16],
des
in
Figur
Moreau
y
y

b
et
ed
terpreted
describ
and
:
to
strategy
ansitive
ecialization
ok
sp
e
the

t
internal
presen
and
briey
CGs
rst
as
e
an
w
This
and
7].
3,
to

b
In
state

w

Language).
bidi-
1
and
2.1
unidirectional
F
oth
r
b
the
.
kno
Unidir
every
e
a

e
variants
form
al
whose
low
of
only
e
one
t,
of
ate
these
1
op
mainly
er
lan-
ators
this
(either
or
in
for
or
us

sp
)
to
but
and
not
sen
b
wn
oth.
in
F

or
on
every
of
nite

vo
automata.

et
abulary

pro
Categorial
and
grammars
for
denitions
every
e
set


wher

is
ontaining
and
the
e
,
elation
a
d

and
ate
F
gorial
ulating
gr
suggests
ammar
analysis
(or
an
CG)
o
is
taking
a
a
nite
e
r
af
elation
e
preserving
and
over
des
RA
el
a
by
to
erimen
in
a
CG
(se
a
1).
transform
simple
to
v
.
een
W
t
e
syn
note
b
w
Let
ho
nor
2
new

neither
in
and

common
to
er
the

assignment

of
e
the
app

es
ate

gory

need
CGs
rst
allo
e
already
w
of
aim,
a
this

h

reac
Finaly
to
strategy
the
ell
element
some
of
e
the
Moreau's
vo
e

nite
abulary
the
o
L
Figure
their
targets
that
.
Grammars
The
(Categories,

Denition
rules

of
and
a

CG
automata
take

the
grammars
form
rom
of
,
two
e
r
strategies.
ewriting
the
schemes:
eexive
h
tr
whic

[11],
of
in
r

wn
and
dene
douille
by
re-
already
F
on
y
schemes.
b
or

lo
osed
stim
prop
a
(F
,
orwar

d

Applic

ation)
b
ositiv
pr
e

p
d,
This
the
v
of
CG
binary-br
sen
tr
lik
e
John
le
or
no
man
ar
fast
assignments
Applic
it
ation)
whose
:
no
examples.
ar
that
lab

le
set
either
lest
but
smal
or
Unidir
p
e
by


CGs
gory
make
e
use
e
of
Example
only
(a
one
CG).
of
ha
these
e
rules
b
(either
used
the
represen
is
natural
or
guage
W
tax,
e
illustrated
)
y
but
example.
not
ex-
of
an
b
result,
oth.
algorithm
The
y
language
where
gener
stands
ate
term
d
oses
(or
for
r
noun,
e
th

pap
o
h.
gnize
ecial