📄 The Formation of Old English Adverbs: Structural Description and Functional Explanation

Summary

Old English adverb formation had received little attention. The only dedicated study, Nicolai (1907), mixes synchronic with diachronic processes and inflection with derivation, does not separate compounding from derivation or words from phrases, and is incomplete. Maíz Villalta revisits the problem with current morphological theory and an exhaustive data set. She extracts every adverb from the Nerthus lexical database, which has about 30,000 headwords drawn mainly from Clark Hall, with Bosworth–Toller and Sweet as secondary sources. Each adverb is then assigned to one formation type: basic (underived), conversion, zero derivation, prefixation, suffixation or compounding. The headline finding is that the adverb is overwhelmingly a derived category. Affixation, especially suffixation with -li:ce, accounts for more than half of all adverbs. The paper also gives a principled rule for deciding the direction of conversion: it runs from the heavily inflected classes toward the nearly invariable adverb. This links adverb suffixes to an old diachronic drift in which inflectional endings become derivational ones. The result is a complete, theory-consistent inventory, with an appendix, that can serve as a reference for Old English word formation.

Key points

• Nerthus contains 1,654 adverbs, about 5% of the lexicon. Nouns make up about 50%, and adjectives and verbs about 20% each. Only 138 adverbs are basic (not derivable by productive means).
• There are 122 converted adverbs. Of these, 98 come from adjectives and 4 from nouns, and 19 have both a noun and an adjective as possible bases. In those 19 cases the adjective is preferred.
• Affixation produces 865 adverbs:
  • Prefixation gives 199 adverbs from 31 prefixes, led by un- (46), for- (25), on- (20) and a:- (14).
  • Suffixation gives 666 adverbs from only 13 suffixes, led by -li:ce (401) and -e (124).
• Several suffixes reflect old case endings: -an (18), -es (46), -a (8) and -um (40). Earlier grammars (Campbell, Lass) treat these as inflection, but the author handles them as derivation. Compounding yields 64 adverbs (about 4%), for example the -mæ:lum series such as æcermæ:lum ‘by acres’.
• The appendix keeps problematic cases in separate classes: 233 hypothetical predicates, 56 bracketing paradoxes (e.g. un- + -li:ce forms) and 64 ambiguous items. Nerthus-style numbered headwords (e.g. a:bu:tan 1 for the adposition, a:bu:tan 2 for the adverb) keep homographs apart.

Interesting ideas

• I1. When a word belongs to two categories without any formal change, the paper takes the more richly inflected class as the source and the less inflected one as the target, giving a reproducible rule for derivation direction.
• I2. Homographs are split into numbered headwords whenever they differ in category, inflectional class or spelling family, so each lexical key is unambiguous.
• I3. Ties between candidate bases are broken by two criteria: which pairing is more frequent overall, and which source shares the target's function (adjective and adverb are both modifiers).
• I4. The analysis does not force uncertain forms into a class. It keeps explicit buckets for hypothetical bases, bracketing paradoxes and genuinely ambiguous items.
• I5. Productivity is compared as the number of derivatives per affix (666 from 13 suffixes vs 199 from 31 prefixes), not just raw counts.

For Mova

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

• I1 (tokens): The en crate could use this rule to link English zero-derived pairs (fast adj/adv, hard, early) to one base lemma, with a recorded reason for the choice.
• I2 (world): The compiled lexicon and ontology could use lemma+index keys, so that knowledge-graph links to WordNet or Wikidata attach to one specific reading rather than a bare string.
• I3 (tokens): Derivational analysis could apply the same two-step tie-breaker when an English -ly form could come from either a noun or an adjective, and log which criterion decided.
• I4 (learn): Silver-data and repair-rule pipelines could keep a labelled "unresolved / paradox" class (e.g. unhappily as un+happily vs unhappy+-ly) instead of guessing, and use it as a negative-control set.
• I5 (tokens): Morphological rules for unknown words could be ranked by this ratio, so the most productive affixes are tried first when segmenting out-of-vocabulary words.