📄 Grover Hudson, Northeast African Semitic: Lexical Comparisons and Analysis

Summary

This is Alessandro Bausi's review of Grover Hudson's 2013 book on the Ethiopian Semitic (ESe) languages. The review appeared in Aethiopica 18. Hudson's book is a large lexicostatistical classification. It measures how closely related languages are by counting shared cognates in a fixed list of basic words. Hudson does not attempt glottochronological dating.

Chapter 1 surveys about thirty language varieties and then narrows the study to fifteen. Chapter 2 builds a 250-item comparison list. Chapter 3 turns the collected forms into an etymological dictionary with proto-language indexes. Chapter 4 presents a dendrogram that splits ESe into five groups. North ESe still forms one group, but South ESe does not come out as a single block.

Bausi praises the data collection and organisation and calls the book a new benchmark. He also raises several objections:

• The background on Geʿez is imprecise.
• Lexicostatistics can tally similarities but cannot judge how much each one matters.
• A dendrogram is not a true genealogical tree.
• Hudson plays down the Agaw substratum on thin lexical evidence, even though that influence is clearly visible in syntax.

Key points

• Issued by Harrassowitz in 2013 as volume 26 of the series Porta Linguarum Orientalium, the book bears the title Northeast African Semitic: Lexical Comparisons and Analysis and runs to viii + 323 pages. It includes 4 maps, 22 tables and 7 tree diagrams.
• Thirty varieties are described, and 15 are analysed. Examples include Tigre, Tigrinya, Geʿez, Gafat, Chaha, Inor, Silt'e, Zay, Harari, Argobba and Amharic, along with their dialects. Dahalik drops out of the comparison tables because data were missing, which leaves 14 languages.
• The word list has 250 items. Bender's 1971 study used 98, although Bender covered all Ethiopian languages.
• The dictionary chapter covers 3,301 words. It indexes Proto-Semitic, Proto-Agaw and Proto-East-Cushitic cognates, and it also gives Proto-ESe reconstructions.
• Bausi notes that the book omits relevant prior work: Kitchen et al. 2009, a Bayesian phylogeny of Semitic, and Fronzaroli's 1961 statistical classification.

Interesting ideas

• I1. Hudson deliberately keeps measuring distance between languages separate from building a time or history story on those distances, because the second step needs extra assumptions.
• I2. Bausi's "counts but does not weigh" criticism shows that tallying overlaps treats every match as equal, so structural evidence can overturn conclusions drawn from vocabulary alone.
• I3. One dataset is published in three views: comparison tables, an etymological dictionary, and indexes keyed by proto-language. Each view serves a different kind of lookup.
• I4. A dendrogram from clustering looks like a family tree, but it only summarises similarity. Bausi warns against reading it as real descent.
• I5. The fixed concept list is larger than earlier ones, and gaps such as missing Dahalik data are stated openly rather than filled in. This keeps the comparisons honest.

For Mova

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

• I1 (learn): Agreement scores between witness parsers should be reported strictly as similarity. They should not be read as proof that the agreed parse is correct without separate negative controls.
• I2 (frames): When frames or coreference decisions are scored, syntactic cues from the UD tree should outweigh mere lexical overlap, in keeping with the no-bag-of-words principle.
• I3 (world): Compiled knowledge such as verb classes and WordNet/Wikidata links could be stored once and exposed through several indexes. Each answer could then be traced back to its single source entry.
• I4 (learn): Clusters of error types or self-induced repair rules should be treated as hypotheses about a shared cause, and checked before being merged into one rule.
• I5 (answer): Evaluation suites should use a fixed item list and report coverage gaps explicitly, so that scores from different readers stay comparable.