📄 Amharic Verbal Types from Text and Lexicon

Summary

The paper describes the verbal morphology of Amharic as it appears in actual newspaper prose. It began as a by-product of an unfinished Amharic–English dictionary meant for foreigners reading the press. Bender, assisted by Negga Shibeshi, counted every word in six daily newspaper issues. He then sorted the 3995 verb occurrences, which come from 400 distinct verbs, using a classification scheme he adapted from Dawkins. The scheme records radical structure, final vowel, gemination type and derivational prefixes. The counts are compared with Mantel-Niecko's 1964 study of a similar sample.

The two studies agree broadly. Verbs with gemination type A make up about half of all occurrences. Basic stems plus the te- and a- derivations cover roughly nine in ten verb tokens. The differences are notable, though: Bender's corpus has more underived stems, far fewer te- forms and many more a- forms. The paper also computes several type-level ratios, such as derived versus basic forms, hypothetical bases, and repeated radicals. These ratios bear on historical questions like where the A/B distinction and quadriliterals come from. Its value is an early, quantitative, corpus-based profile of a Semitic verb system.

Key points

• The source journal (Folia Orientalia XV) is dated 1974 in the scan, not 2026. The corpus is three issues each of Addis Zemen and Ye-Ityoppya Dimts, yielding 3995 verb forms of 400 verbs. Mantel-Niecko's comparison sample had 5000 verbs and 430 roots.
• Share of occurrences by gemination type:
  • type A: about 1/2
  • type B: about 1/5
  • type C and "hollow" verbs: about 1/7 each
• Basic, T and A forms together cover 91% of occurrences in this corpus and 87% in Mantel-Niecko's. Basic stems make up 66% here versus 54% there.
• About 38% of all verb occurrences are basic forms. On average, 18% of occurrences are built on hypothetical bases, rising to 47% for quadriliterals.
• Repeated radicals are about five times more common in type IA than in IB (0.05 vs 0.01), and 30% of quadriliterals show a 1212 pattern. The commonest verb here, serra "do, make, work", is absent from Mantel-Niecko's top ten.

Interesting ideas

• I1. The paper deliberately reruns an earlier count on fresh text and reports a second percentage without nouns and other categories, so the two studies share a denominator; this makes cross-study differences interpretable rather than artefacts of scope. — For Mova (learn): When comparing a parser or tagger against an older baseline, Mova could always report metrics recomputed on the baseline's exact label scope alongside its own full-scope figures.
• I2. Each verb receives a compact code built from independent dimensions (radical class, vowel, gemination pattern, special markers), so one label summarises several orthogonal properties that can later be aggregated separately. — For Mova (tokens): The English morphology crate could tag lemmas with composite, decomposable inflection-class codes, for example irregular-past pattern plus spelling-change type, so lemma errors can be grouped by class.
• I3. Derivational prefixes are treated as operators stacked in a fixed order onto a stem, which turns a combined form into an explicit, reproducible derivation history. — For Mova (frames): Voice and causative alternations, such as passive, causative "make X do" and reflexive, could be modelled as ordered operators on a base frame, keeping the role remapping explainable step by step.
• I4. A root that is never seen bare but only inside derived forms is still posited as a "hypothetical base" and counted, separating abstract lexical units from what is attested. — For Mova (world): The compiled verb-class lexicon could hold unattested base entries, for instance for English words seen only as derivatives, flagged as inferred so the knowledge graph stays honest about evidence.
• I5. Counts of distinct verbs and counts of their occurrences are kept side by side, which shows that a small class can dominate running text. — For Mova (learn): Silver-data quality checks should report both type and token coverage per verb class, so rare-but-frequent and common-but-sparse classes are not confused.