1. Constraint Grammar ESSLLI
Wednesday:
Syntax

2. QA NER What is CG used for? VISL grammar games: Machinese parsers
News feed and relevance filtering
Opinion mining in blogs
Science publication monitoring QA
Machine translation
Spell- and Grammar checking
Corpus annotation
Relational dictionaries: DeepDict
NER
Annotated corpora:
CorpusEye

3. CG Input remember CG is modular!
output from one module becomes input to the next module
pre-
processor
analyzer
nomorph CG
your CG
text
output
cat file | preprocessor | analyzer | vislcg3 -g engcg.nomorph.cg | vislcg3 -g yourcgfile ...
echo ”text” | preprocessor | analyzer | vislcg3 -g engcg.nomorph.cg | vislcg3 -g yourcgfile ...

4. CG input it can be very simple preprocessor:
perl -wnpe 's/ /\n/g; s/([^\w\n])/\n$1\n/g; s/\n\n+\ng/;'
analyzer:
#!/usr/bin/perl -w
use utf8;
open(FH, "< /home/eckhard/parsers/eng/lex/baselex.eng");
while (<FH>) {
s/\t\/\n\t/g;
($word,$tags) = split /,/;
$lex{$word} .= "\t" . $tags; }
while (<>) {
$word =$_;
chop $word;
print "\"<$word>\"\n";
if ($lex{$word}) {
print $lex{$word};

5. Syntactic annotation styles
Focus on syntactic form
Phrase structure grammar (PSG) -> labelled brackets
Dependency grammar (DG) -> labelled arcs
Focus on syntactic function
Constraint grammar (CG) -> dependency pointers
Focus on semantic function
Case roles (Filmore)
Lexical Functional Grammar (LFG) CG

6. Syntactic models 1. The flat classical model: word function, no form
My fat cat never eats fish S A V O
word-based
psychologically easy to grasp
function markers attached to semantically heavy words
easy to turn into tags:
My determiner PRE-N
fat adjective PRE-N
cat noun S
never adverb A
eats verb V
fish noun O CG

7. 2. Dependency grammar strictly token based (e.g. Tesnière)
expresses syntactic form as asymmetrical relations (“arcs”) between head tokens and dependent tokens
no zero tokens, no nonterminal nodes
each dependent is allowed 1 head (exc. secondary arcs)
directed acyclic graphs
projective or non-projective (crossing branches / discontinuity)
My fat cat never eats fish. CG

8. Dependency grammar annotation
The #1->3
last #2->3
sattelites #3->9
launched #4->3
by #5->4
the #6->7
US #7->5
never #8->9
reached #9->0 ROOT
orbit #10->9
The <id=1> <head=3>
last <id=2> <head=3>
sattelites <id=3> <head=9>
launched <id=4> <head=3>
by <id=5> <head=4>
the <id=6> <head=7>
US <id=8> <head=5>
enver <id=8> <head=9>
reached <id=9> <head=0> ROOT
orbit <id=10><head=9> CG

9. Dependency grammar as trees
reached
diagnóstico
never
orbit
The
last
sattelites
launched by US CG
the

10. Dependency grammar with brackets “a la PSG”, e.g. TIGER
Penn-style:
[V eats [N cat [DET my][ADJ fat]] [ADV never] [N fish]]
Vertical:
[V eats
[N cat
[DET my]
[ADJ fat] ]
[ADV never]
[N fish] CG

11. 3. Constituent Grammar
hierarchical word grouping with non-terminals (e.g. Chomsky)
syntactic form, no (or implicit) function
expressed by rewriting rules, where a nont-terminal node is rewritten as a sequence of non-terminals and terminals (words or word classes) s -> np vp np -> art n vp -> v np
Pure Constituent Grammar: CG

12. Classical PSG with phrase labels
(VP) NP
ADV |
never V |
reached N |
orbit
ART |
The
ADJ |
last N |
sattelites
PCL
V-PCP |
launched PP
PRP | by NP
ART |
the
PROP | US CG

13. PSG annotation Penn Treebank bracketing: Labeling opening brackets
[NP A minha irmã] [VP não fala [PP com [NP as amigas]]]
SUSANNE Treebank bracketing: Labeling all brackets (cf. EAGLES)
[NP A minha irmã NP] [VP não fala [PP com [NP as amigas NP] PP] VP]
Vertical indented (her with part of speech on one line):
[NP [Art A] [Det minha] [N irmã] NP] [VP [Adv não] [V fala] [PP [Prp com] [NP [Art as] [N amigas] NP] PP] VP] CG

14. Adding function:
Dependency Grammar with function: adding function (“edge labelse”) to dependency arcs
reached S Od fA
sattelites
never
orbit DN DN DN
The
last
launched
PASS by DP US DN CG
the

15. Constituent Grammar with function:
NEGRA, TIGER: cat labels (mother) vs. edge label (daughter CG

16. Constituent Grammar with function:
VISL (function:form labels for each node)
Vertical Notation:
STA:cl
S:np
=DN:art O
=H:n governo
=DN:prop Cardoso
P:v-fin
A:pp
=H:prp com
=DP:np
==DN:art a
==H:n crise
Graphical Notation: CG

17. CG

18. CG

19. 4. Constraint Grammar (CG)
CG as a descriptive paradigm
function-first approach with token-based function tags
Classic CG: shallow depedency (attachment direction, head type)
depth and constituents only implicitly marked
My @>N (pointer to head type: N)
fat @>N
cat @SUBJ> (direction pointer without head type)
never @ADVL>
eats @FMV (top node)
fish @<ACC CG

20. Adding full numbered dependency
The <artd> ART #1->3
last #2->3
sattelites N P #3->9
launched V PCP2 #4->3
by #5->4
the <artd> ART #6->7
US PROP @P< #7->5
never #8->9
reached V #9->0
orbit N S #10->9
$. #11->0 CG

21. 2. Interactive syntactic tree buildingCG

22. CG

23. BEFORE / AFTER SECTIONS
BEFORE-SECTIONS
AFTER-SECTIONS
Run only once (cp. VISLCG's MAPPING / CORRECTIONS)
Especially for
adding ambiguity, e.g. Syntactic, semantic roles etc.
Post-processing errors from previous modules

24. Mapping (MAP, ADD) MAP (@SUBJ) TARGET (N) IF (NOT *-1 NON-PRE-N)
MAP (N) (NOT *-1 NON-PRE-N)
Usually as a special section (MAPPING or BEFORE-SECTIONS), but in cg3 allowed anywhere
Strictly ordered
Both MAP and ADD can be used to add tags, but:
MAP "closes" a line for further mapping (but not SUBSTITUTE!) even if the mapped tags do not contain the flagged prefix
ADD maps, but allows further mapping
MAPed tags can be "seen" by later mapping rules, even in the same section

25. Substitutions SUBSTITUTE (S/P) (P) TARGET (@<SC)
IF + (P) BARRIER CLB)
Replaces a tag or tag chain with another, useful for:
correcting input from other modules, e.g. stochastic taggers
correcting lower level CG once higher lever information is available
spell or grammar checkers
Usually as a special section (CORRECTIONS or BEFORE-SECTIONS), but in cg3 allowed anywhere
'TARGET' and 'IF' are optional
Strictly ordered
SUBSTITUTE does not "close" a line for mapping
SUBSTITUTEd tags can be "seen" by later SUBSTITUTE or Mapping rules, even in the same section

26. CG Contexts
Context conditon: word form “<...>”, base form “....”, tag A-Z, <[a-z]> @[A-Z], combinations ...
direction: + (right), - (left)
Position marker:
0 self
local right: 1, 2, 3 ..., local left: -1, -2, -3, ...
Globality
* continue until match is found
** continue also across context match to fulfil further (linked) conditions
0* nearest neighbour: search in both directions
Careful: C, e.g. *1C (only unambiguous readings)

27. CG contexts 2 NOT: conditions can be negated (NOT *1 VFIN)
contexts can be LINKed
(*1C xxx LINK 0 yyy LINK *1 zzz)
searches can have a BARRIER
(*1 N BARRIER VFIN)
contexts can be ANDed
IF (0 xxx) (*1 yyy) (NOT *-1 zzz)
contexts can be negated as a whole
(NEGATE *1 ART LINK 1 ADJ LINK 1 N)

28. CBARRIER (**1 N CBARRIER VFIN) ; like BARRIER, but only if unambiguous
i.e. less strict than BARRIER

29. NEGATE (NEGATE *1 (AUX) LINK 1 (@AUX<)) ;
(NEGATE *-1 N LINK -1 DEF) ;
implements aspects of the TEMPLATE idea (being able to refer to - and to negate - chunks of internally linked tokens
will invert the result of the entire LINK'ed chain that follows
whereas NOT will only invert the result of the immediately following test

30. Unification
set names (LIST, SET) can be defined as variables by prefixing $$
in a CG rule, all occurrences of the $$ set will be unified to mean the same set member
LIST ROLE = %AG %PAT %TH %LOC ;
SELECT $$ROLE (-1 KC) (-2C $$ROLE) ; # (4-in-1 rule)
the $$ occurrence in the target position is the primary one (i.e. the one the others unify with)
if $$ only is used in contexts, add KEEPORDER to force a safe interpretation of the first occurrence as the primary one:
REMOVE KEEPORDER (NEGATE 0 $$CASE LINK -1 N + $$CASE)