Arabic Synonyms Extraction using Singular Value Decomposition and Self-Organizing Maps

Abstract

Arabic is a morphologically rich language where many words share the same root. This causes a well-documented problem when using automatic synonym extraction, where standard methods based on embeddings and cosine similarity -- the Baseline -- consistently return root-related word forms instead of true functional synonyms. This thesis addresses this problem. This thesis proposes a three-stage pipeline combining a pre-trained Arabic language model (BERT), SVD, and the SOM. BERT generates contextual word representations; SVD reduces the embedding dimensionality of any given word to suppress noise; and the SOM reorganises the space so that functionally equivalent words cluster together rather than root-related ones. The SVD-SOM pipeline was trained on approximately 3,800 Arabic articles covering 157 target words from variety of domains including education, economics, law, and media. A key limitation of the SVD-SOM pipeline is the absence of a standard Arabic synonym evaluation benchmark, which makes performance measurement inherently difficult. Evaluation was therefore conducted by comparing the SVD-SOM pipeline output against the Microsoft Word Arabic Thesaurus and several established Arabic online dictionaries (Almaany, Al-Maajim, Arabic Wiktionary), and the system was shown to retrieve true functional synonyms directly and reliably. To make this comparison quantitative, a lexically-grounded weighted synonym evaluation (the WSS@k and wDCG@k measures, defined in Chapter 2) scores each retrieved candidate on a 0--1 closeness scale and rewards ranking true synonyms near the top of the list. The results are clear. For the word تعليم (Education), the Baseline ranked terms such as أستاذ (Professor) and طلابي (Student-related) within its top-15 output -- neither of which is a functional synonym. The proposed SVD-SOM pipeline ranked تدريس (Teaching) and تثقيف (Enlightenment) in the top two positions, exactly matching the Microsoft Word Arabic Thesaurus. This pattern was reproduced consistently across the case studies presented in Chapter 5. A quantitative cross-validation using an architecturally independent Sentence-BERT model across all 2,512 experimental scenarios showed a modest but consistent improvement of the SVD-SOM pipeline over the Baseline. Future work should prioritise building a publicly available Arabic synonym benchmark and extending the SVD-SOM pipeline to include Arabic dialects and downstream applications such as information retrieval.