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

Abdelkader El Mahdaouy, Saïd El Alaoui Ouatik, Éric Gaussier

TL;DR

This paper introduces a novel hybrid method for automatically extracting Multi-Word Terms (MWTs) from Arabic text. By combining linguistic analysis with a new statistical measure called NLC-value, which incorporates contextual information alongside termhood and unithood, the method significantly improves MWT extraction accuracy compared to existing approaches.

What Problem Does This Paper Address?

Extracting Multi-Word Terms (MWTs) – like "artificial intelligence" or "climate change" – is crucial for many Natural Language Processing (NLP) tasks, including information retrieval and text categorization. While robust methods exist for languages like English, Arabic presents unique challenges due to its rich morphology and variations in word forms. Existing MWT extraction techniques often fall into three categories: purely linguistic, purely statistical, or hybrid. However, purely statistical methods struggle with low-frequency MWTs, and purely linguistic methods can be language-dependent and inflexible. Hybrid methods aim to combine the strengths of both, but many existing approaches for Arabic either don't fully leverage contextual information or don't adequately handle MWT variations. This paper addresses this gap by proposing a hybrid approach that specifically tackles MWT variation and introduces a new statistical measure that integrates contextual nuances.

Methods

The proposed method employs a two-stage hybrid approach:

  1. Linguistic Filter: This stage uses a Part-of-Speech (POS) tagger (specifically, the AMIRA toolkit trained on the Penn Arabic TreeBank) to identify candidate MWTs based on syntactic patterns. Crucially, this filter also handles MWT variations, including graphical (orthographic errors), inflectional (different word forms for gender/number), morpho-syntactic, and syntactic variants. A normalization method and light stemming algorithm are applied to address these variations.
  2. Statistical Filter: After linguistic filtering, candidate MWTs are ranked using statistical measures. The paper introduces a novel statistical measure, the NLC-value, which combines:
  3. Termhood: The degree to which a linguistic unit is related to a specific domain concept (measured by C-value).
  4. Unithood: The degree of strength or stability of collocations (measured by LLR - Log-Likelihood Ratio, a more accurate measure than T-score used in prior work).
  5. Contextual Information: The terminological status of the context surrounding a candidate MWT (measured by N-value).

The NLC-value is calculated as a weighted combination of the LC-value (C-value with LLR-based unithood) and the N-value. This approach aims to overcome the limitations of measures that only consider termhood or unithood.

Datasets

The authors built a new corpus specialized for the environmental domain. This corpus consists of 1666 files with 53,569 tokens extracted from the website "Al-Khat Alakhdar," covering topics like pollution, noise effects, water purification, and climate change. For evaluation, they also utilized the AGROVOC thesaurus and the IATE database as reference lists.

Key Insights and Ideas

  • Context is King (for MWTs): Integrating contextual information into statistical measures significantly boosts MWT extraction performance, as demonstrated by the NLC-value outperforming measures that ignore context.
  • Arabic MWTs Need Special Handling: The inherent variations in Arabic (graphical, inflectional, etc.) are a significant hurdle. Addressing these variations at the linguistic filtering stage is crucial for accurate MWT extraction.
  • Hybrid is Better: Combining linguistic and statistical approaches effectively mitigates the weaknesses of each individual method, leading to more robust MWT extraction.
  • LLR for Unithood: Using the Log-Likelihood Ratio (LLR) for unithood provides a more accurate assessment of word adhesion within a term compared to simpler measures like the T-score.

Key Results

The paper's proposed NLC-value measure consistently outperformed other statistical measures (LLR, C-value, NC-value, NTC-value, LLR+C-value) in terms of precision for both bi-grams and tri-grams. For instance, when considering the top 100 MWTs, the NLC-value achieved 82.0% precision, compared to 80.0% for NTC-value and 75.0% for LLR. Similar improvements were observed for the top 200 and 300 MWTs. The results suggest that the combination of contextual information with termhood and unithood is essential for high-quality MWT extraction.

Limitations and Future Work

The authors acknowledge that the evaluation was performed on a domain-specific corpus (environmental). While they built this corpus to mimic properties of existing domain-specific corpora, its scope is limited. Future work includes using the extracted MWTs in an information retrieval system, as complex terms can provide a better representation of document content than single words. They also plan to explore further improvements to the statistical measures and potentially expand the linguistic analysis.

Summary updated

Abstract

Automatic Multi-Word Term (MWT) extraction is a very important issue to many applications, such as information retrieval, question answering, and text categorization. Although many methods have been used for MWT extraction in English and other European languages, few studies have been applied to Arabic. In this paper, we propose a novel, hybrid method which combines linguistic and statistical approaches for Arabic Multi-Word Term extraction. The main contribution of our method is to consider contextual information and both termhood and unithood for association measures at the statistical filtering step. In addition, our technique takes into account the problem of MWT variation in the linguistic filtering step. The performance of the proposed statistical measure (NLC-value) is evaluated using an Arabic environment corpus by comparing it with some existing competitors. Experimental results show that our NLC-value measure outperforms the other ones in term of precision for both bi-grams and tri-grams.
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