Automatic Tagging of Arabic Text: From Raw Text to Base Phrase Chunks

Mona T. Diab, Kadri Hacioglu, Daniel JurafskyNAACL 2004 · OpenReview

Abstract

To date, there are no fully automated systems addressing the community's need for fundamental language processing tools for Arabic text. In this paper, we present a Support Vector Machine (SVM) based approach to automatically tokenize (segmenting off clitics), part-of-speech (POS) tag and annotate base phrases (BPs) in Arabic text. We adapt highly accurate tools that have been developed for English text and apply them to Arabic text. Using standard evaluation metrics, we report that the SVM-TOK tokenizer achieves an Fβ=1 score of 99.12, the SVM-POS tagger achieves an accuracy of 95.49%, and the SVM-BP chunker yields an Fβ=1 score of 92.08.

Introduction

Arabic presents unique challenges for Natural Language Processing (NLP) due to its rich, templatic morphology. A major hurdle is the presence of clitics—short function words like prepositions, conjunctions, and pronouns that physically attach to the beginning (proclitics) or end (enclitics) of a word stem. For example, the English phrase "and by their virtues" is written as a single fused word in Arabic, as illustrated below.

Morphological structure of an Arabic word showing clitics and stem.

At the time of this research, no fully automated fundamental NLP tools existed for Arabic. Researchers relied on rule-based morphological analyzers that required manual intervention to resolve ambiguities. This paper introduces a fully automated pipeline to segment clitics (tokenization), assign syntactic categories (POS tagging), and group words into non-recursive phrases (BP chunking).

Methodology

The pipeline treats all three tasks as supervised classification problems using Support Vector Machines (SVMs). A key design choice was the use of language-independent features, adapting techniques originally developed for English NLP:

  • Word Tokenization (SVM-TOK): Tokenization is framed as a character-level classification task. Using a sliding window of 11 characters, the model tags each character with labels (e.g., Beginning of Proclitic, Inside Stem) to segment clitics from the stem.
  • POS Tagging (SVM-POS): This is modeled as a token-level classification task using a collapsed version of the Arabic TreeBank tagset (24 tags). The model utilizes a 5-token window and features such as character n-grams and token types.
  • Base Phrase Chunking (SVM-BP): Modeled as a 1-of-19 classification task using the standard IOB (Inside, Outside, Beginning) tagging scheme to identify 9 phrase types.

Datasets and Evaluation

The primary dataset used was the Arabic Penn TreeBank 1 (v2.0), consisting of Agence France Presse (AFP) newswire articles. The data was transliterated into Latin-based ASCII characters using the Buckwalter scheme. The corpus includes roughly 140k words, split into 4,000 sentences for training and 400 for testing.

Key Results

The SVM-based pipeline achieved performance comparable to state-of-the-art English NLP tools. The SVM-TOK tokenizer, in particular, demonstrated the superiority of data-driven methods over traditional rule-based systems.

Comparison of tokenization performance between SVM-TOK and rule-based systems.
  • Tokenization: SVM-TOK achieved a near-perfect $F_{\beta=1}$ score of 99.12, significantly outperforming a dictionary-assisted rule-based baseline (93.71).
  • POS Tagging: SVM-POS reached an accuracy of 95.49%, surpassing the 92.2% baseline. For comparison, an English SVM POS tagger trained on similar data volumes typically achieves around 94.97%.
  • BP Chunking: SVM-BP yielded an $F_{\beta=1}$ score of 92.08, approaching the performance of English chunkers (93.48) on comparable data.

Discussion and Conclusion

The most significant takeaway is that standard SVM setups designed for Indo-European languages work remarkably well on Arabic when the problem is framed correctly (e.g., character-level tagging for clitics). The data-driven approach proved more robust than dictionaries, which suffer from coverage issues.

While POS tagging is slightly bottlenecked by linguistic ambiguities in Arabic—specifically the overlap between nouns and adjectives—the overall pipeline provides a highly accurate, automated foundation for Arabic NLP. Future work will focus on expanding training data and extending the pipeline to handle semantic role labeling.

Summary updated

Abstract

To date, there are no fully automated systems addressing the community's need for fundamental language processing tools for Arabic text. In this paper, we present a Support Vector Machine (SVM) based approach to automatically tokenize (segmenting off clitics), part-of-speech (POS) tag and annotate base phrases (BPs) in Arabic text. We adapt highly accurate tools that have been developed for English text and apply them to Arabic text. Using standard evaluation metrics, we report that the SVM-TOK tokenizer achieves an Fβ=1 score of 99.12, the SVM-POS tagger achieves an accuracy of 95.49%, and the SVM-BP chunker yields an Fβ=1 score of 92.08.
BibTeX
@inproceedings{diab2004automatic,
  author = {Mona T. Diab and Kadri Hacioglu and Daniel Jurafsky},
  title = {{Automatic Tagging of Arabic Text: From Raw Text to Base Phrase Chunks}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2004},
  url = {https://openreview.net/forum?id=B1XKu7bOWH},
}
Figure