📄 Mengistu Amberber: Verb Classes and Transitivity in Amharic

Summary

This is not an empirical NLP paper. It is a short book review by Olga Kapeliuk of Mengistu Amberber's Verb Classes and Transitivity in Amharic, published in 2002 as Lincom Studies in Afro-Asiatic Linguistics 6. The PDF header dates the review to Aethiopica 10 (2007), not 2012. Amberber's book groups Amharic verbs by their derivational prefixes and by valency, meaning how many nominal participants a verb takes. Where verbs break the expected pattern, he explains the exceptions through lexical semantics, drawing on Levin's English verb classes.

Kapeliuk praises the book's compactness and theoretical fluency. Her main criticism is that it imposes English-shaped categories on Amharic:

• It splits passive from reflexive even though both use the same prefix.
• It counts a two-clause causative construction as part of the verb system.
• It sorts intransitives into unaccusative and unergative using contestable semantic intuitions.
• It ignores the historical origins of look-alike stems.
• Several of its example sentences are mistranslated.

The review is useful as a warning about transferring an English-based lexical classification to another language without checking it against data.

Key points

• The book has 108 pages and six chapters:
  • Chapter 1 covers basic unprefixed verb morphology.
  • Chapters 2–5 cover verbs derived with the prefixes tä-, a- and as-, alone and in combination.
  • Chapter 6 covers compound forms built with the verb alä.
• The book's label set is adapted from Levin's English Verb Classes (1993). It divides intransitives into "unaccusatives" (states) and "unergatives" (actions).
• The reviewer finds some class assignments hard to justify:
  • 'exit' is classed as unaccusative but 'migrate' as unergative.
  • 'grow' (of plants) is classed as a state but 'laugh' as an action.
• She argues that some bound/independent stem pairs, such as 'fear' vs. 'fructify', are historical homophones with separate Ge'ez sources. On that view they are not related derivations.
• She identifies two mistranslated examples:
  • One is glossed as passive ('was returned') but can only mean 'returned'.
  • In the 'sweep' example, the roles are reversed. A bare instrumental, without the preposition b-, must be indefinite and stand next to the verb.

Interesting ideas

• I1. Classifying verbs by valency alone can split forms that share one morpheme, as with passive vs. reflexive tä-, so form-based and valency-based labels should be stored separately and not merged into one class.
• I2. A semantic rule that predicts morphological behaviour is only trustworthy if it is checked across the whole lexicon, not on a handful of chosen verbs. Unlike phonology, the lexicon is too heterogeneous for broad generalizations.
• I3. A periphrastic "make someone do X" construction should be analysed as a link between two clauses, not as a single verb form, unless the language really binds the two verbs tightly.
• I4. Look-alike stems may be accidental homophones with different histories, and forcing a synchronic derivational link between them creates false structure.
• I5. Example glosses can silently swap semantic roles, as in the broom-and-doorway case, so constructed examples need a native-reading check of word order and definiteness.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): keep the morphological construction and the argument-structure frame as separate, explainable fields, so a frame decision never silently overrides what the surface form says.
• I2 (world): when compiling verb-class knowledge from VerbNet or Levin-style lists, run each class-to-behaviour rule over the full verb inventory and report exception counts before trusting it.
• I3 (frames): model causatives as a causing frame embedding a caused frame. This keeps who-did-what-to-whom explicit across clauses instead of flattening it into one predicate.
• I4 (tokens): when lemmatizing or linking derived forms, require evidence beyond surface similarity, so unrelated homographs are not merged under one lemma.
• I5 (learn): add negative controls in which role-swapped readings of a sentence must be rejected. This catches models or teachers that only look plausible on schematic examples.