Constraint Grammar

En français

Constraint Grammar is a tool that can be used to POS-tag ambiguous text. There are free constraint grammars developed outside the Apertium project for: Norwegian (the Oslo-Bergen tagger), Sámi languages (from Giellatekno), Faroese (also from Giellatekno), Finnish (by Fred Karlsson).

Terminology

See also: Apertium stream format

• cohort — a surface form of a word, along with its analyses (possible lexical units), an ambiguous lexical unit.

Apertium equivalent: ^words/word<n><pl>/word<vblex><pres><p3><sg>$

• baseform — the lemma of a word.
• reading — a single analysis of a word.

Apertium equivalent: ^word<n><pl>$

• wordform — a surface form of a word.

Basic Rule Format

Sets

Sets are defined like this:

LIST VERB = vblex vbser ;     # matches <vblex> or <vbser>
LIST NSG = (n sg) ;           # matches <n><sg>
LIST TO = "to" ;              # matches the lemma "to"
LIST CASE = (n nom) (n acc) ; # matches <n><nom> or <n><acc>

Context Patterns

Context patterns look like this:

([LOCATION][MODIFIERS] [PATTERN])

PATTERN can be a lemma, set of tags, or the name of a set.

Symbol	Meaning
0	The current word
1	The word following the current word
-1	The word preceding the current word
2	The word 2 words after the current word
C	Every reading this position must match the pattern (normally only 1 has to)
*	In that position or further in that direction

(0 (v))      # the current word must have a verb reading
(1 VERB)     # the following word matches the set "VERB"
(-1 "to")    # the previous word must be "to"
(2C (n))     # every reading of the word after the next one must be a noun
(1* (pr))    # the current word has a preposition after it
(-2* (pron)) # there is a pronoun at least two words before the current word

Rules

Rules look like this:

SELECT [FORM] IF [CONTEXT] ;
REMOVE [FORM] IF [CONTEXT] ;

Where FORM is a lemma, set of tags, or the name of a set and CONTEXT is a set of patterns.

SELECT VERB IF (1 (det)) ;    # prefer reading from set "VERB" if following word has <det> tag
REMOVE (n) IF (-1 (adv)) ;    # disprefer <n> reading if preceding word has <adv> tag
SELECT (p1) IF (-1* (prn p1)) # prefer 1st person reading for verbs
               (0 (v)) ;      # preceded at any distance by a 1st person pronoun

Note on parenthesis

The use of parentheses to distinguish between tags and lists/sets seems to be the main confusing point for people learning CG. If we have the morphological tags tag1 and tag2, then we can have rules like this:

LIST set1 = tag1 ;
LIST set2 = (tag1 tag2) ; # matches a word with both tag1 and tag2
LIST set3 = tag1 tag2 ;   # matches a word with tag1 or tag2
LIST word = "hello" ;

SELECT:1a (tag1) (1 word) ;
SELECT:1b  set1  (1 word) ;   # equivalent to 1a

SELECT:2a (tag1 tag2) (1 word) ;
SELECT:2b  set2       (1 word) ;   # equivalent to 2a

SELECT:3a tag1 (1 word) ;
SELECT:3b tag2 (1 word) ;
SELECT:3c set3 (1 word) ;   # equivalent to 3a and 3b combined

SELECT:1c  set1  (1 ("hello")) ; # equivalent to 1a (or 1b)

Languages using CG in Apertium

• Breton
• Welsh
• Norwegian Nynorsk and Norwegian Bokmål
• Sámi languages
• Irish Gaelic

and many others. The following languages currently (2014-06-27) have CG's of over 100 rules:

• 3888 apertium-nno (based on the Oslo-Bergen tagger)
• 3649 apertium-sme (from Giellatekno)
• 2275 apertium-nob (based on the Oslo-Bergen tagger)
• 1552 apertium-est
• 1524 apertium-fin (based on Fred Karlsson's)
• 850 apertium-dan
• 594 apertium-gle 1207 in gle-eng.rlx
• 453 apertium-fao (from Giellatekno)
• 298 apertium-spa
• 279 apertium-bre [1]
• 255 apertium-cat
• 205 apertium-hbs [2], also hbs-mkd.rlx with syntax rules
• 190 apertium-isl
• 131 apertium-cym [3]
• 76 apertium-tur
• 127 apertium-eng
• 118 apertium-mkd
• 123 apertium-tat [4]
• 308 apertium-rus [5]
• 150 apertium-kaz

When is CG needed?

Currently some of the CG rules written in the above language pairs may be written as forbid rules in the TSX format used by apertium-tagger. If the rules for your language pair can be written in the .tsx format, you can go for an easier design without a CG module in that language pair.

Editor support

• CG-3 IDE – the official vislcg3 CG IDE
• Gedit syntax highlighting (also for any other editor that uses gtksourceview)
• Emacs emacs mode for editing and testing CG grammars (highlighting + IDE-like features)

See also

• Apertium and Constraint Grammar -- installation and use
• Introduksjon til føringsgrammatikk -- a HOWTO, in Norwegian bokmål
• Rule-based finite-state disambiguation -- GsoC 2012 project by User:Krvoje, a "CG light" (or, a more apertiummy CG) with rules in XML compiled to an FST
• Constraint Grammar/Speed – some tips on speeding up your rules
• Constraint Grammar/Optimisation – ideas on how to optimise the vislcg3 engine

External links

• VISL CG-3 Development Information + documentation and downloads
• Basic Tutorial for VISL CG-3
• cg-mode for emacs, gives basic syntax highlighting and indentation
• Kevin Donnelly's CG tutorial
• Hulden M, Francom J (2012) Boosting Statistical Tagger Accuracy with Simple Rule-Based Grammars, Proc. LREC 2012, p. 2114-2117 shows how 20 hours (very little time!) writing disambiguation rules gives substantial improvements. Some of the rules shown may also be implemented in the TSX format used by apertium-tagger.