📄 Sisay Fissaha Adafre: Adding Amharic to a unification-based machine translation system – an experiment

Summary

This item is not a research article. It is a German-language book review by Rainer Voigt (Freie Universität Berlin) in Aethiopica 8, pp. 257–264. The reviewed work is Sisay Fissaha Adafre's 2004 Saarland dissertation. That thesis builds a formal morphosyntactic description of Amharic for the unification-based CAT2 machine translation system, with morphology implemented in the Xerox finite state tools. Voigt, a Semitist, explicitly leaves the MT engineering unassessed and examines only the linguistic analysis.

He credits the author for rejecting weak parts of autosegmental (McCarthy-style) templatic theory, notably the obligatory contour principle, and for rejecting Baye Yimam's reductions of quadriliteral roots. However, he finds the Amharic data itself badly flawed:

• gemination is frequently missing or misplaced;
• person suffixes are wrong;
• grammatical glosses are incomplete;
• cleft sentences and the copula are misanalysed;
• the "collocation" chapter groups unrelated phrase types together.

His verdict is that whatever CAT2's merits, input data of this quality cannot support Amharic translation. For library users, the review is a cautionary case: grammar resources for low-resource languages need expert auditing.

Key points

• The reviewed book was published by Peter Lang in 2004 as volume 5 of the Sabest series (xxii + 208 pp.). It targets the CAT2 MT system, with morphology built in Xerox finite state tools.
• The morphology chapter (pp. 13–74) critiques the templatic analysis. Voigt argues that a CVC:VC template, which treats gemination as length, avoids the ad hoc reassociation and spreading rules that CVCCVC requires.
• The author rightly rejects the obligatory contour principle by citing roots like qql and ddb. Voigt adds a list of further qq/dd-initial roots drawn from Kane's Amharic–English dictionary (1990).
• Rejecting Baye Yimam's proposal to drop a predictable second radical is justified because the reduction causes collisions. For example, dfrs would collapse into the existing root drs ("arrive"), and snbt into sbt.
• Chapter 6 (pp. 187–193) proposes bi-gram and tri-gram collocations as lexical units. Voigt objects that it mixes compounds, adverbials and fully compositional phrases. He notes that qalä mïlïllïs ("interview") is really a compound noun, and that qädäm sïl ("earlier") is compositional rather than idiomatic.

Interesting ideas

• I1. A proposed universal morphological constraint is refuted by systematically listing counterexample roots from a dictionary, rather than by arguing in the abstract.
• I2. Encoding gemination as a length feature on one slot, instead of two copied consonant slots, removes a whole cascade of corrective rules.
• I3. Multiword "collocations" found by frequency should be split into types (true compounds, idioms, compositional phrases) before any of them becomes a lexical unit.
• I4. In the cleft example, a masculine object-suffix reading clashes with a feminine name, which exposes the misanalysis; agreement mismatch thus works as a cheap validity check.
• I5. Errors in transcription and glossing, repeated across a resource, are judged fatal for the downstream system regardless of the formalism's quality.

For Mova

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

• I1 (learn): Mova's self-induced repair rules could be checked against lexicon-derived counterexample lists as a negative control before a rule is accepted.
• I2 (tokens): Analogous phenomena in the English morphology crate (doubled consonants in "stopped", "running") could be modelled as features, keeping lemma rules short and explainable.
• I3 (world): Only non-compositional expressions should enter the compiled lexicon or ontology, while compositional ones stay parsed, so frames remain derivable.
• I4 (frames): Gender and number agreement between a pronoun-like element and its candidate antecedent could automatically flag implausible coreference or role assignments.
• I5 (learn): Gold and silver treebanks should get targeted expert audits of systematic errors, since such errors propagate into every trained parser.