arXiv preprint · August 2026
Qaamuuska-NLP: A Structured 46K-Entry Somali Lexicon Extracted from a Print Dictionary
1Ogaal Labs, Somalia
2Chungbuk National University, South Korea
Abstract
Somali NLP still has relatively few structured lexical resources that preserve the linguistic information found in traditional dictionaries. We present Qaamuuska-NLP, a machine-readable reconstruction of Qaamuuska Af-Soomaaliga, the 2012 monolingual Somali dictionary by Puglielli and Mansuur. The resource contains 46,314 dictionary records extracted directly from the PDF's native text layer using a deterministic, rule-based pipeline, without OCR or machine-learning-based extraction. Where available in the source, records include part of speech, noun gender, verb class and transitivity, plural information, subject domains, synonym references, cross-references, homonym indices, and numbered definitions. The original entry text is retained alongside the parsed representation to support inspection against the source.
We describe the extraction pipeline, the structure and coverage of the resulting resource, and the main classes of residual parsing errors. We also examine the distribution of grammatical and lexical information in the dictionary and use a simple surface-form prediction probe to test how strongly selected morphological labels are reflected in Somali word forms. Qaamuuska-NLP complements existing Somali morphological, lemmatization, corpus, and multi-dictionary lexical resources by preserving the structure of a single major monolingual dictionary in a reproducible computational form.
Keywords · Somali NLP · computational lexicography · lexical resources · dictionary digitization · morphology · low-resource languages
The source dictionary
Qaamuuska Af-Soomaaliga (Dictionary of the Somali Language), edited by Annarita Puglielli and Cabdalla Cumar Mansuur and published by Roma Tre Press in 2012, is a major monolingual reference dictionary for Somali. Its entries contain more than headwords and definitions: they record grammatical categories, noun gender, verbal information, plural forms, semantic distinctions, subject labels, synonyms, homonyms, and cross-references.
The dictionary is distributed as a fixed two-column PDF rather than a documented fielded dataset. Its text can be extracted digitally, but the lexical structure is expressed through typography, abbreviations, entry ordering, superscript markers, punctuation, and recurring conventions. Plain text extraction therefore does not directly provide records that can be searched by grammatical field, linked through cross-references, or used systematically in NLP pipelines.
The extraction processes PDF pages 24–916 with pdfplumber, working from the
native text layer rather than rendered page images.
Extraction pipeline
-
Column-aware text extraction
Each page is divided at its horizontal midpoint and the two columns are extracted independently, left before right, so the intended reading order survives. Running headers are removed with source-specific heuristics and the result is appended to a continuous buffer, so entries spanning a page break are recovered intact.
-
Entry segmentation
Records are located using a recurring entry-start pattern: a headword, an optional superscript homonym marker, then one of the dictionary's recognised part-of-speech codes. The span between one entry start and the next becomes a candidate record body.
-
Field parsing
Each candidate is decomposed into headword, homonym index, original POS code, normalized POS category, noun gender, verb conjugation class and transitivity, plural information, parenthetical verbal information, subject-domain labels, numbered definitions, synonym references and cross-references. Numbered definitions are kept separate rather than stored as one undifferentiated string.
-
Cross-reference resolution
A second pass builds an index over the extracted headwords and connects each redirect record — signalled in the source by markers such as
ldandeeg— to the integer identifier of its target, using the homonym marker where the reference carries one.
Coverage of principal extracted fields
| Field / property | Records | % of all |
|---|---|---|
| Total records | 46,314 | 100.0 |
| Definition(s) | 25,240 | 54.5 |
Redirect (ld / eeg) | 21,032 | 45.4 |
| Redirect resolved to target ID | 20,075 | 43.3 |
| Gender annotation | 34,570 | 74.6 |
| Verb-class annotation | 11,444 | 24.7 |
| Synonym reference(s) | 11,415 | 24.6 |
| Homonym index | 9,212 | 19.9 |
| Noun plural metadata | 4,698 | 10.1 |
| Subject-domain label | 1,945 | 4.2 |
Predicative-only (khabar) | 529 | 1.1 |
Part-of-speech distribution
| Part of speech | Records | % |
|---|---|---|
| Noun | 34,726 | 75.0 |
| Verb | 11,445 | 24.7 |
| Exclamation | 71 | 0.2 |
| Pronoun | 54 | 0.1 |
| Particle | 11 | < 0.1 |
| Preposition | 5 | < 0.1 |
| Numeral | 2 | < 0.1 |
| Total | 46,314 | 100.0 |
Among the 34,726 noun records, 34,570 carry a gender annotation (99.6% of nouns): 18,410 feminine, 13,934 masculine, and 2,226 marked as permitting either. Of the 11,445 verb records, 11,444 carry a conjugation class — 4,170 Class I, 3,067 Class II, 2,193 Class III and 2,014 Class IV — alongside 6,528 intransitive, 4,821 transitive and 96 bitransitive markings.
Subject domains
| Domain | Source label | Records |
|---|---|---|
| Medicine | daaw. | 360 |
| Physics | fiis. | 239 |
| Mathematics | xis. | 226 |
| Biology | baay. | 220 |
| Chemistry | kiim. | 219 |
| Religion | dii. | 171 |
| Geography | juqr. | 152 |
| Geology | jool. | 106 |
| Botany | bot. | 104 |
| Music | muus. | 39 |
| Politics | siyaa. | 36 |
| History | taar. | 35 |
| Commerce | dhaq. | 14 |
| Zoology | c.nafl | 13 |
| Law | qaan. | 8 |
| Psychology | c.naf | 3 |
| Total | 1,945 |
Medicine, physics, mathematics, biology and chemistry together contribute 1,264 entries, roughly 65% of all domain-labelled records. This distribution reflects the source dictionary's coverage of specialized terminology and should not be read as the topical distribution of Somali vocabulary as a whole.
Surface-form predictability of morphological labels
As a diagnostic, a suffix back-off classifier — trained only on word-final character sequences of one to six characters, under an 80/20 headword-disjoint split — was asked to predict labels from the surface form alone. It receives no definition, dictionary metadata, contextual sentence, external corpus, or pretrained representation.
| Task | Records | Majority acc. | Suffix acc. | Macro-F1 |
|---|---|---|---|---|
| Verb class (I–IV) | 11,444 | 0.378 | 0.935 | 0.936 |
| Noun gender (m/f/b) | 34,570 | 0.538 | 0.864 | 0.874 |
| POS (noun/verb/other) | 46,314 | 0.746 | 0.869 | 0.560 |
Verb conjugation class and noun gender are strongly predictable from word-final morphology, which is consistent with the role of suffixation in Somali. Coarse part of speech behaves differently: the classifier improves accuracy from 0.746 to 0.869, but its macro-F1 is only 0.560, because nouns account for about 75% of records and the remaining categories represent a very small fraction of the lexicon.
The experiment does not establish the accuracy of a Somali morphological analyzer on independent running text. It is a within-resource surface-form predictability analysis, not an extrinsic downstream evaluation.
Limitations
- The resource derives from a single dictionary. Its coverage, sense distinctions, terminology, dialectal choices and grammatical conventions reflect the editorial decisions of Qaamuuska Af-Soomaaliga rather than a balanced sample of contemporary Somali usage.
- The pipeline is deliberately source-specific. Its rules depend on this dictionary's two-column layout, abbreviations, typography and entry patterns.
- Characters are preserved without a general Unicode normalization pass. More than one apostrophe code point occurs in the source, which affects exact headword matching, duplicate detection and cross-reference resolution. This explorer folds those code points at query time, so both spellings find the same records.
- Residual artifacts remain and are reported rather than hidden: 957 redirects (4.6%) do not
resolve, the
(headword, homonym_index)pair collides for 88 records, and small fractions of definitions are empty, unusually short, or retain text resembling the start of an adjacent record. - Annotations are monolingual and source-specific; they have not been mapped to Universal Dependencies or exported through TEI Lex-0 or OntoLex-Lemon.
Ethics and data availability
Qaamuuska-NLP is derived from Qaamuuska Af-Soomaaliga, a copyrighted monolingual Somali dictionary published by Roma Tre Press. The source dictionary is publicly accessible as a PDF, but public access does not by itself imply permission to redistribute its lexical content in a new structured form.
Permission to publicly redistribute the complete extracted lexical dataset is pending. The released artifacts are the extraction code, aggregate statistics, and a small illustrative sample of records, at github.com/Ogaal-Labs/somali_dictionary_lexicon. The extraction code is MIT-licensed; that license applies only to the software and does not extend to the lexical content of the source dictionary.
The resource is intended for research and educational work in Somali language technology. It contains lexicographic material and no personal data, human-subject annotations, or information collected from individual participants.
Citation
Abdullahi Mohamud Ahmed and Abdulkadir Ugas. Qaamuuska-NLP: A Structured 46K-Entry Somali Lexicon Extracted from a Print Dictionary. arXiv preprint, August 2026.