A Study of Association Measures and their Combination for Arabic MWT Extraction

Abdelkader El Mahdaouy, Saïd El Alaoui Ouatik, Éric Gaussier

Automatic extraction of Multi-Word Terms (MWTs)—such as "climate change" or "renewable energy"—is a foundational task for domain-specific knowledge management, translation, and information retrieval. While research has matured for European languages, Arabic presents unique challenges due to its rich morphology and structural variations. This paper proposes a novel hybrid method for Arabic MWT extraction that integrates linguistic patterns with a new statistical measure called NLC-value.

The Challenge of Arabic MWT Extraction

Arabic presents three primary hurdles for automated extraction systems:

  1. Morphological Complexity: A single word often contains prefixes and suffixes that alter its grammatical role, complicating simple pattern matching.
  2. Structural Variation: Concepts can be expressed through multiple syntactic structures (e.g., Noun-Noun vs. Noun-Adjective), requiring flexible linguistic filters.
  3. Data Sparsity: Technical terms often appear with low frequency in specialized corpora, causing purely statistical methods—which rely on high-frequency counts—to fail.

Methodology

The authors employ a two-stage hybrid approach consisting of a linguistic filter followed by a statistical ranking filter.

1. Linguistic Filter

This stage identifies candidate terms by processing the raw text through several layers:

  • POS Tagging: Utilizing the AMIRA toolkit to assign parts of speech.
  • Syntactic Pattern Matching: Identifying sequences such as Noun + Noun, Noun + Adjective, or Noun + Preposition + Noun.
  • Variation Handling: The system identifies graphical (spelling), inflectional (gender/number), and morpho-syntactic variants to ensure different forms of the same term are grouped together.
Flowchart of the Arabic MWT extraction system architecture.

2. Statistical Filter (NLC-value)

To rank the identified candidates, the authors introduced the NLC-value, a weighted measure that balances three critical aspects of terminology:

  • Unithood: The statistical strength of the bond between words, measured here using the Log-Likelihood Ratio (LLR).
  • Termhood: The degree to which a phrase is a technical term, based on a modification of the classic C-value.
  • Context (Nvalue): A measure of the term's environment; candidates appearing near other known technical terms receive higher scores.

The final formula is defined as: NLC-value(a) = 0.8 · LC-value(a) + 0.2 · Nvalue(a).

Experimental Setup

Due to the lack of a standard Arabic MWT benchmark in the environmental domain, the authors developed a custom corpus:

  • Environmental Corpus: 1,666 files (approximately 53,569 tokens) sourced from environmental news sites.
  • Evaluation: Results were validated against established multilingual thesauri, including AGROVOC and IATE.

Key Results and Insights

The proposed NLC-value consistently outperformed existing state-of-the-art measures, including LLR, C-value, and NTC-value.

Precision comparison of various statistical measures for MWT extraction.

As shown in the experimental data, the NLC-value achieved a precision of 82.0% for the top 100 candidates, significantly higher than the 71.0% achieved by the standard C-value. Even as the candidate list expanded to the top 300, the NLC-value maintained a robust precision of 73.0%, whereas the next best hybrid method (NTC-value) dropped to 69.7%.

Key Takeaways:

  • LLR Superiority: The Log-Likelihood Ratio proved more effective for measuring word stability in Arabic than the T-score typically used in European models.
  • The Importance of Pre-processing: Linguistic normalization (e.g., standardizing Alif and Ya variants) is essential before applying statistical measures to Arabic text.
  • Holistic Integration: Success in extraction requires balancing termhood, unithood, and context simultaneously.

Conclusion and Future Work

While the NLC-value shows significant promise, the authors note that the current evaluation was limited to a relatively small corpus (~53k tokens). Future research will focus on testing the system on larger, more diverse datasets and integrating the extraction tool into full-scale Information Retrieval systems to measure its impact on search accuracy for Arabic users.

Summary updated

BibTeX
@misc{mahdaouy2014study,
  author = {Abdelkader El Mahdaouy and Saïd El Alaoui Ouatik and Éric Gaussier},
  title = {{A Study of Association Measures and their Combination for Arabic MWT Extraction}},
  year = {2014},
  eprint = {1409.3005},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1409.3005},
}
Figure