Abstract
We investigate the adequate unit of analysis for Arabic Mention Detection (MD), experimenting with various segmentation schemes and feature sets. Our results demonstrate that while morphologically segmented data is superior in resource-constrained environments, Arabic TreeBank (ATB) style segmentation performs better when rich features are available. We also demonstrate that combining these segmentation schemes leads to further performance gains.
The Challenge of Arabic Morphology
Mention Detection (MD) involves identifying named, nominal, and pronominal references. In Arabic, this task is complicated by complex morphology where a single word can represent multiple linguistic units. For example, a single token might include a conjunction, a preposition, and a definite article attached to a stem.
Treating these "clumps" as single units leads to data sparsity, while over-segmenting can lose semantic context. This study seeks the optimal "Goldilocks" level of segmentation for Arabic NLP tasks.
Methodology
The researchers evaluated four segmentation schemes using a Maximum Entropy (MaxEnt) classifier:
- Words: Simple punctuation-based separation.
- Arabic TreeBank (ATB): Splits affixes only when they function as independent phrasal constituents (e.g., separating "and" but keeping "the" with the noun).
- Morphological (Morph): Aggressively splits all prefixes and suffixes.
- Character-level: Treats each letter as a token.
Performance was tested across four feature categories: Lexical (n-grams), Stem (word stems), Syntactic (POS tags/parsing), and Semantic (outputs from external ACE-trained classifiers).
Key Results
The experiments revealed a clear trade-off between sparsity reduction and contextual preservation.
As shown in the table above:
- Morphological segmentation is highly effective at reducing Out-of-Vocabulary (OOV) issues. With only lexical features, it outperformed the word-level approach by nearly 8 F-measure points (74.1 vs 66.4).
- ATB segmentation excels when rich semantic features are included, reaching an F-measure of 79.0. It better preserves the local context necessary for identifying long-span mentions.
- Character-level models are insufficient for Arabic MD, performing poorly (around 22 F-measure) compared to units that carry semantic weight.
The Hybrid Peak
The study's most significant finding is that these schemes are complementary. By using the output of a morphological model as a feature for an ATB-based model, the system achieves a "hybrid peak" of 79.1 F-measure. This approach leverages the robustness of morphological splitting while maintaining the structural advantages of ATB segmentation. These improvements were confirmed to be statistically significant ($p < 0.02$).
Conclusion
For optimal Arabic Mention Detection, the choice of unit depends on available resources. While morphological segmentation is a powerful "sparsity killer" for basic models, the highest performance is achieved by integrating multiple linguistic views. Future work should focus on deeper architectural integration of these segmentation strategies rather than relying on one as a simple feature for the other.
Summary updated