📄 Formal analysis of some aspects of Amharic noun phrases

Summary

Amharic, a Semitic language of Ethiopia with rich morphology, had almost no formal syntax in unification-based grammars. This makes it hard to build syntax-driven machine translation for it. The paper is part of an effort to add Amharic to the transfer-based CAT2 MT system, and it focuses on one fragment of the grammar: noun phrases, especially the definite article.

The author first describes the data:

• Definiteness is marked by a suffix (-u), while demonstratives are separate words placed before the noun.
• Definiteness and accusative marking can move onto prenominal modifiers such as adjectives.

Neither the pure DP analysis nor the pure NP analysis fits these facts well, and the earlier movement-based account (Girma's KP analysis) is criticised. Instead, the author treats the article as an affix, encoded as a feature of the word that carries it, following Wintner. Netter's intermediate projections and Streiter's analysis strategy complete the grammar, which is written in CAT2. The author argues that this lexical approach explains the data better than transformational accounts. It also simplifies the language-neutral interface structure used for transfer.

Key points

• Prior work: the only earlier syntactic work on Amharic was a prototype PCFG (Atelach, 2003). Earlier CAT2 work by the same author covered only the lexicon and morphology.
• Core facts about definiteness:
  • The suffix -u cannot co-occur with a demonstrative on the same noun.
  • When an adjective comes between them, the adjective may take -u, giving over-determination.
  • Possessive suffixes also exclude -u on the noun, but an adjective modifying a possessed noun takes it.
• Accusative marker: the direct object marker -n follows the definite suffix. It attaches to every definite prenominal modifier, and on an indefinite NP it yields a generic reading.
• Critique of Girma (1994): it wrongly groups prepositions with accusative -n, and it treats the article as an independent syntactic head without good motivation.
• Encoding in CAT2: definiteness is a feature (type=def) inside an extended head feature (ehead). This feature is transferred through the Interface Structure, and the target-language generator decides how to realise it.

Interesting ideas

• I1. Bound function morphemes like articles are modelled as features of their host word rather than as separate syntactic nodes, which keeps one representation across languages that write them as words or affixes.
• I2. Agreement information may sit on modifiers, not only on the head noun, so features must be allowed to rise from several constituents of the phrase.
• I3. Mutual-exclusion facts (demonstrative vs. article, possessive vs. article) are used as hard tests that any candidate analysis must pass.
• I4. Transfer goes through an abstract structure that stores definiteness only as a semantic feature, and each target language decides its surface form.
• I5. Definiteness is separated from the indefinite vs. generic readings of bare nouns.

For Mova

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

• I1 (tokens): morphology could emit definiteness/case as features on the host token, so clitic-like English forms ('s, n't) are handled uniformly before parsing.
• I2 (tree): when building NP structure, definiteness and possession cues should be collected from determiners and modifiers, not read off the head noun alone.
• I3 (learn): such co-occurrence bans can serve as negative-control checks on silver trees, flagging parses that give one NP two determiners.
• I4 (back): the meaning graph should keep definiteness as an entity attribute and let generation choose "the", "this", a possessive, or nothing.
• I5 (world): entities from bare or generic NPs could be flagged as kinds rather than individuals, so they are not added as concrete objects in the world of the text.