The Effect of Automatic Tokenization, Vocalization, Stemming, and POS Tagging on Arabic Dependency Parsing

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 10–18, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics The Effect of Automatic Tokenization, Vocalization, Stemming, and POS Tagging on Arabic Dependency Parsing Emad Mohamed Suez Canal University Suez, Egypt emohamed@umail.iu.edu Abstract We use an automatic pipeline of word tokenization, stemming, POS tagging, and vocalization to perform real-world Arabic dependency parsing. In spite of the high accuracy on the modules, the very few errors in tokenization, which reaches an accuracy of 99.34%, lead to a drop of more than 10% in parsing, indicating that no high quality dependency parsing of Arabic, and possibly other morphologically rich languages, can be reached without (semi-)perfect tokenization. The other module components, stemming, vocalization, and part of speech tagging, do not have the same profound effect on the dependency parsing process. 1. Introduction Arabic is a morphologically rich language in which words may be composed of several tokens and hold several syntactic relations. We define word to be a whitespace delimited unit and token to be (part of) a word that has a syntactic function. For example, the word wsytzwjhA (ﺎﻬﺟﻭﺰﺘﻴﺳﻭ)(English: And he will marry her) consists of 4 tokens: a conjunction w, a future marker s, a verb inflected for the singular masculine in the perfective form ytzwj, and a feminine singular 3rd person object pronoun. Parsing such a word requires tokenization, and performing dependency parsing in the tradition of the CoNLL-X (Buchholz and Marsi, 2006) and CoNLL 2007 shared task (Nivre et al, 2007) also requires part of speech tagging, lemmatization, linguistic features, and vocalization, all of which were in the human annotated gold standard form in the shared task. The current study aims at measuring the effect of a pipeline of non gold standard tokenization, lemmatization, vocalization, linguistic features and POS tagging on the quality of Arabic dependency parsing. We only assume that we have gold standard sentence boundaries since we do not agree with the sentence boundaries in the data, and introducing our own will have a complicating effect on evaluation. The CoNLL shared tasks of 2006 and 2007 used gold standard components in all fields, which is not realistic for Arabic, or for any other language. For Arabic and other morphologically rich languages, it may be more unrealistic than it is for English, for example, since the CoNLL 2007 Arabic dataset has tokens, rather than white space delimited words, as entries. A single word may have more than one syntactically functional token. Dependency parsing has been selected in belief that it is more suitable for Arabic than constituent-based parsing. All grammatical relations in Arabic are binary asymmetrical relations that exist between the tokens of a sentence. According to Jonathan Owens (1997: 52): “In general the Arabic notion of dependency and that defined in certain modern versions e.g. Tesniere (1959) rest on common principles”. With a tokenization accuracy of 99.34%, a POS tagging accuracy of 96.39%, and with the absence of linguistic features and the use of …

Content truncated to 500 words