PENN ARABIC TREEBANK

The Penn Arabic Treebank project started with the assumption that the
existing methodological principles and accumulated experience and 'wisdom'
of the previous Penn treebanks could and should be very helpful in
jumpstarting the Arabic Treebank. A team of three (Mohamed Maamouri, Ann
Bies, and Hubert Jin) started work in November 2001 with the objective of
completing the Treebank annotation of an initial 100K corpus of AFP
newswire.

The study and analysis of Modern Standard Arabic syntax and of the
appropriateness and adaptability of the Penn English Treebank methodology
to the Arabic language by Mohamed Maamouri and Ann Bies surprisingly showed
that existing Penn English Treebank methodology and tags worked
satisfactorily for the main Arabic syntactic issues and could therefore be
used with few changes.  For the most part, our syntactic/predicate-argument
annotation of newswire Arabic follows the bracketing guidelines for the
Penn English Treebank where possible.  Basic sentence structure, node
labels and functional tags, arguments/adjuncts, coordination, and empty
categories are treated in essentially the same way.  Draft guidelines can
be found in guidelines-TB-1-28-03.pdf.

Some points where the Arabic Treebank differs from the Penn English Treebank: 

* Arabic subjects are analyzed as VP internal, following the verb.  The
subject (labeled NP-SBJ) is inside VP after verb, and is frequently
pro-drop (NP-SBJ * ).  If the lexical subject precedes the verb, it is
labeled NP-TPC (topicalized) and traced to (NP-SBJ *T* ) following the
verb. 

* Matrix clause (S) coordination is possible and frequent. 

* The function of NP objects of transitive verbs is directly shown as NP-OBJ. 

* Co-reference is shown always on the node label, never on the empty
category token itself. 

* Gapping co-reference is always shown as '=' indexing, for both the
template and the subsequent gap filling items.


CLITICIZATION 

The prevalence of cliticization in Arabic sentences of determiners,
prepositions, conjunctions, and pronouns led to a necessary difference in
tokenization between the POS files and the treebanking files.  Clitics that
play a role in the syntactic structure are split off into separate tokens
(e.g., object pronouns cliticized to verbs, subject pronouns cliticized to
complementizers, cliticized prepositions, etc.).  Clitics that do not
affect the structure are not separated (e.g., determiners).  There is a
file containing the rules guiding the separation of the clitics:
appendix/bin/func-subst2.list

   
SAMPLE BRACKETED TREE

(S (CONJ [w]) 
   (VP (IV3MS+VERB_IMPERFECT [yblg]) 
       (NP-SBJ (NOUN [Edd]) 
		(NP (DET+NOUN+NSUFF_MASC_PL_ACCGEN [Alm$rdyn]))) 
	(PP-LOC (PREP [fy]) 
		(NP (NOUN [kwntyp]) 
		    (NP (NOUN_PROP [lws]) (NOUN_PROP[Anjlys])))) 
	(PP (PREP [nHw]) 
	    (NP	(NP (QP (NUM [48]) (NOUN [Alf])) 
		    (NOUN [$xS]))
		(SBAR (WHNP-1 0) 
		      (S (PP-PRD (PREP [byn]) 
				 (NP (NP (PRON_3MP [hm])) 
				     (NP-1 *T*))) 
			  (NP-SBJ (QP (NOUN+NSUFF_FEM_SG [tsEp]) 
				      (NOUN [AlAf])) 
				  (NOUN [Tfl])))))) 
	(SBAR-TMP (CONJ [bynmA]) 
		  (S (VP (PRT (NEG_PART [lA]))
			 (IV3FS+VERB_IMPERFECT [ttsE]) 
			 (NP-SBJ (DET+NOUN [AlmrAkz]) 
				 (DET+NOUN+NSUFF_FEM_SG [AlxASp]))
			 (PP (PREP [l]) 
			     (NP (NOUN [{stqbAl]) (POSS_PRON_3MP [hm]))))
			 (PP (ADV [swY]) (PREP [l]) 
			     (NP (NUM [00631]) (NOUN [sryr])))))))
	(PUNC [.]))


TAGLIST

Constituent tags:

S	sentence
NP	noun phrase
VP	verb phrase
PP	prepositional phrase
SBAR	S-bar (subordinate clause, complementizer or WH- and sentence)
SBARQ	S-bar that is a question
SQ	S that is a question
NX	noun head in certain complex coordination contexts
PRN	parenthetical
PRT	particle
QP	quantity phrase (multi-word numbers)
ADJP	adjective phrase
ADVP	adverb phrase
FRAG	fragment
WHNP	WH- noun phrase
WHPP	WH- prepositional phrase
WHADJP	WH- adjective phrase
WHADVP	WH- adverb phrase
CONJP	conjunction phrase (multi-word conjunction)
INTJ	interjection
NAC	Not-A-Constituent (mostly rightward moved conjuncts with conjunction)
UCP	Unlike-Coordinated-Phrase (dominates coordination of NP and PP, e.g.)
X	unknown, technical problem, etc.


Function (dash) tags:

-SBJ	subject
-OBJ	object (ONLY BRAND NEW TAG FOR ARABIC!)
-TPC	topicalized
-PRD	predicate
-PRP	purpose
-CLR	CLosely-Related (non dative PP argument of verb, for the most part)
-LOC	locative
-DIR	directional
-MNR	manner
-TMP	temporal
-ADV	adverbial (for NP, S and SBAR)
-LGS	LoGical Subject (NP object of by PP in passives)
-NOM	nominal (for S and SBAR)
-DTV	dative
-VOC	vocative
-BNF	benefactive
-EXT	extent
-CLF	cleft (as in, it-cleft)
-HLN	headline
-TTL	title


TREEBANK PARSING ANNOTATORS:

Wigdan EL MEKKI
Mohamed MANSOUR
Tasneem GHANDOUR
Ichraf AMGHOUZ
Niama LAADIOUI