The Word Based Parsing Technique (part-5)

4.5. PoS tagging

PoS tagging of The Arabic sentence and word disambiguation, was previously addressed indirectly by developing language resources for inputing to statistical disambiguators [22] [63], or directly either by rule based or statistical based system. [64] [34][65][59][57][45][66][67][54][68][69]

For statistical disambiguation, it was important to have the minimum adequate number of PoS tags that can be useful for a statistical algorithm to train on limited manually disambiguated corpora. The number of unique PoS tags extracted in the word generation phase, previously described in this chapter, is above 4700 tags. We selected only 3033 on which the developed rules are applied. The use of this number of PoS tags with pure statistical model (such as Diab’s [45]), needs huge manually annotated corpus (And it is not reported in the literature that such corpus exists, as of this writing).

Another problem exists in front of statistical disambiguation tools, is the long distance dependency. As reported by Diab [45], the window length used in PoS tagging of AMIRA I is 5 tokens, which will fail in the following sentence:

Sentence #4.1:

* طلب المحامي من المستشارين بعد الاطلاع على المستندات مد مدة الدفاع

PoS Tagging and BP chunking output is:

* [VP طلب/VBD] [NP المحامي/NN] [PP من/IN المستشارين/NNS] [PP بعد/IN الاطلاع/NN] [PP على/IN المستندات/NNS] [PP مد/IN مدة/NN الدفاع/NN]

It can be seen that the object of the sentence (مد) is marked as (IN) which denotes a preposition, while the word itself has no solution with BAMA mentioning it as preposition. The output of BAMA for this word is:

LOOK-UP WORD: مد SOLUTION 1: (مَدَّ) [mad~-u_1] مَدّ/VERB_PERFECT+ـَ/PVSUFF_SUBJ:3MS      (GLOSS):  + extend/stretch/spread out + he/it <verb>  SOLUTION 2: (مَدّ) [mad~_1] مَدّ/NOUN      (GLOSS):  + extension/lengthening/spreading +

Listing 4.1: BAMA output for the word (مد)

AMIRA is trained on Penn Arabic Treebank [22], which used BAMA as its morphological analyzer. This proves that systems such as AMIRA may produce results that are not in the domain of the word analysis.

The proposed system produces correct analysis taking into consideration the long distance dependency. Figure 4.2 is one of the analyses produced with the proposed system.

Figure 4.2: Morpho-syntactic analysis and PoS tagging of Sentence #5 with DLG

It can be seen in Figure 4.2 that the correct object is linked to the verb with link “On” (nominal object). And the object itself is marked as “n” noun. It can be also seen that all of the other words are assigned valid PoS tags. The following table summarizes the resulted PoS tags for this sentence:

طلب	pv-pvss-3ms	Perfect-verb+Perfect-Verb-Suffix-3rd person-masc.-singular
المحامي	d-n	Determiner + Noun
من	prep-0	Preposition
المستشارين	d-n-ns-md-ag	Determinter+Noun+Noun-Suffix-Masc-Dual-accusative/genitive form
بعد	prep-n	Preposition (can be noun)
الاطلاع	d-n	Determinter+Noun
على	prep-0	Preoposition
المستندات	d-adj-ns-fp	Determinter+Adjective+Noun-Suffix-Feminine-Plural
مد	n	Noun
مدة	n-ns-fs	Noun+Noun-Suffix-Feminine-Singular
الدفاع	d-n	Determinter+Noun

Table 4.5: PoS description for sentence #5

As depicted in Figure 4.2 and Table 4.5, the developed parsing system for DLG, produces with every solution the PoS tagging for every given word.