📄 The recursive formation of Old English non-verbal categories. Productivity and constraints

Summary

The paper describes how Old English builds nouns and adjectives by stacking derivational affixes on bases that are already derived, which it calls recursivity. It asks which affix combinations occur, what limits them, and why speakers would derive a derivative again. The data are all prefixed and suffixed nouns and adjectives in the Nerthus lexical database of Old English. Each derivation is modelled step by step with Pounder's (2000) Process and Paradigm Model, attaching one affix per operation. Recursion depth is split into two levels: two affixes on a plain base, and three affixes where the base is itself recursive. The author finds that suffixes do almost all of the stacking. The most frequent suffixes, -nes for nouns and -lic for adjectives, only ever appear in the outermost position. The motivation also differs by category. Stacking on nouns mainly yields abstract concepts, a semantic effect. Stacking on adjectives mostly adds expressive near-synonyms, a pragmatic effect. This gives a quantitative picture of the limits of derivational productivity in a historical lexicon.

Key points

• Data: the Nerthus database (consulted September 2014) supplies 4,370 derived nouns and 3,218 derived adjectives. Of these, 2,001 are prefixed and 5,587 are suffixed.
• Recursive forms: 388 words are recursive, made up of 257 nouns and 131 adjectives.
  • 377 have two affixes.
  • Only 11 have three, e.g. ealdordōmlicnes 'authority', built as EALDOR → ealdordōm → ealdordōmlic → ealdordōmlicnes.
• Suffix combinations: recursive suffixation gives 58 distinct combinations for nouns and 38 for adjectives.
  • -nes has the most tokens among noun suffixes.
  • -ð pairs with the widest set of inner (prefinal) suffixes.
  • -lic dominates adjectives both in number of types and in variety of patterns.
• Productivity rule: strict positional limits signal productivity when they restrict the inner slot. They signal low productivity when they restrict the outer slot. Prefixes such as un- and healf- show no positional limits and can be inner or outer.
• Prefix functions: recursive prefixation mainly intensifies (ðurh-), mitigates (healf-, under-) or negates (un-, and-). The account fits Lieber's (2004) Redundancy Restriction.

Interesting ideas

• I1. Each derived word is analysed as a chain with one affix added per step, and words are graded by depth: simplex, one affix, two affixes, or three affixes. This makes the internal history of complex words countable and comparable.
• I2. Some suffixes are only ever found at the outer edge of a word, the way -nes and -lic close off derivation. Such slot restrictions work as hard structural constraints on possible analyses.
• I3. An affix's productivity is judged from two signals together: how many lexemes carry it, and how many distinct combination patterns it enters. Two signals are more informative than raw frequency alone.
• I4. Each morphological operation is written as a record holding form change, meaning function and category change, with its restrictions attached. One rule format can then cover affixation, conversion and compounding.
• I5. The paper separates stacking that creates new content (abstract nouns) from stacking that mostly adds expressiveness (adjectival near-synonyms). Stacked forms therefore do not always carry new meaning.

For Mova

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

• I1 (tokens): the morphology module in crate en could output an explicit derivation chain (e.g. carelessness → careless → care) with a depth value, so the lemma and base decisions stay explainable.
• I2 (tokens): segmentation could reject any analysis that places closing suffixes like English -ness or adverbial -ly inside another suffix, pruning wrong splits of unknown words.
• I3 (learn): when inducing derivational rules from silver data, affixes would be ranked on both type count and pattern variety before rules may analyse unseen words.
• I4 (world): derivation rules in the compiled knowledge could use this record, so a noun formed from an adjective is linked to the property node of its base rather than treated as unrelated.
• I5 (world): adjectives with redundant extra suffixes could map to the same property as their base, with a style or intensity flag, while abstract nouns get their own concept entity.