|
چکیده (انگلیسی)
|
scarce resources for Classical and specialized domains, poor understanding of domain sensitivity, Arabic morphological segmentation underlies nearly all downstream NLP tasks, yet progress has been limited by and the absence of standardized benchmarks. This thesis addresses these gaps through a unified, data-centric research program spanning resources, modeling, evaluation, and infrastructure.
We introduce Noor-Ghateh, a manually annotated ۲۰۴,۶۳۰ token corpus of Classical Jurisprudential Arabic with fine-grained segmentation, part-of-speech tags, lemmas, roots, and morphological features, validated through a rigorous quality-assurance protocol. A tri-domain diagnostic evaluation of Farasa, CAMeL Tools, and ALP shows that performance differences across Modern Standard, Scriptural, and Jurisprudential Arabic are not a uniform degradation, and localizes errors to stacked proclitics and annotation-convention mismatches. We then present Noor-Seg, a reproducible, FAIR-aligned segmentation framework achieving ۹۹.۶۷% in-domain boundary-level accuracy (۹۸.۸۹% word-level exact match; ۸۸.۴۷% on word types unseen in training) with transfer to Quranic Arabic (۸۹.۶%) but not to Modern Standard Arabic (۶۱.۰%). We further propose ABWS, a unified Arabic segmentation benchmark, introducing a canonical boundary representation and a linguistically weighted diagnostic metric (Critical Boundary Accuracy), and showing how annotation conventions affect cross-system comparison. Finally, we ask what segmentation boundaries are worth to downstream systems. Controlled part-of-speech tagging experiments on the ۲۰۴,۶۳۰-token corpus show that a single-word character-level BiLSTM (five-fold cross-validation) reaches ۹۹.۷۳% accuracy on raw text and ۹۹.۷۷% with explicit boundary marking, so no reliable improvement from boundary marking was established for that architecture, whereas the word-level BiLSTM is sensitive to boundary deletion. In an equal-budget boundary-deletion analysis with ۵,۴۰۰ exposures per condition, the net damage to word-level predictions was ۱۷.۱۹%, ۳۲.۰۴% and ۲۱.۸۱% for PREFIX->PREFIX, PREFIX->STEM and STEM->SUFFIX deletions, against ۰.۳۹%, ۳.۴۸% and -۳.۸۱% for the character-level model; the downstream cost of boundary errors is therefore architecture- and boundary-type-dependent. Across six controlled extrinsic settings, Noor-Ghateh-derived dictionary segmentation reduced vocabulary sparsity or out-of-vocabulary rates where these properties were measured, without producing reliable downstream improvements in the six evaluated settings, clarifying the conditions under which morphological segmentation resources transfer to practical NLP systems.
Together, these contributios establish Arabic morphological segmentation as a reproducible, data-centric, and diagnostically principled discipline. |
|
|
|
|
|
