📄 The syntax and semantics interface of present-day and Old English speech verbs : say and tell versus secgan and tellan

Summary

The paper compares how meaning is mapped onto syntax for two English speech verbs, say and tell, and their Old English ancestors secgan and tellan. It works within the Functional-Lexematic Model (FLM) and borrows logical structures, Aktionsart classes and Actor/Undergoer macroroles from Role and Reference Grammar (RRG). Each verb domain gets one "lexical template", a single representation that combines a semantic decomposition with argument slots. Say/secgan get an accomplishment template, and tell/tellan get a causative accomplishment template. The individual syntactic alternations are then derived by reducing that maximal template. Old English examples come from the Dictionary of Old English Corpus, checked against Bosworth–Toller. Present-day patterns come from Levin (1993) and Mairal Usón (1993). The alternations are ranked by how closely they match a prototypical transitive clause. By that ranking, modern say and tell are more transitive than their Old English counterparts but cover less semantic space in the speech domain. This shows a diachronic shift in how verbs link syntax and semantics.

Key points

• Old English data come from The Dictionary of Old English Corpus, which holds 3037 texts (poetry, prose, glosses, glossaries, inscriptions). The findings were verified against two reference works: Toller and Campbell (1972) and Bosworth and Toller (1973).
• The say/secgan template is a speaker activity followed by the hearer becoming aware of the content. Tell/tellan uses the same pieces linked by CAUSE.
• Each template has internal variables (content α, addressee β, language γ) and external, syntactically realised variables (x speaker, y hearer, z content). Alternations arise from which internal variables are promoted to external argument positions.
• The Lexical Template Modeling Process derives alternations by four kinds of reduction from the maximal template:
  • suppressing external variables,
  • instantiating internal ones,
  • dropping operators,
  • fusing with other templates.
• Speak/sprecan show the reverse trend: Old English sprecan allows more transitive patterns than modern speak. The Lexical Iconicity Principle (Beta Reading) is used to infer that the modern verbs lost semantic coverage.

Interesting ideas

• I1. One maximal template per verb class, with every observed frame derived by a small set of reduction operations, replaces a hand-listed inventory of subcategorisation frames.
• I2. Separating internal variables (ontological constants of the class, such as content, addressee and language) from external argument slots lets a role exist in meaning even when it is absent from the tree.
• I3. Arguments are ranked from fully prototypical transitive (two referential arguments) through clausal and coordinated complements down to activity uses, rather than treating transitivity as yes/no.
• I4. Macroroles plus the Actor–Undergoer hierarchy make role assignment a default computed from the logical structure instead of a value stored per verb.
• I5. Syntactic coverage is used as evidence for a lexeme's semantic breadth within its domain, so counting alternations doubles as a semantic measurement.

For Mova

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

• I1 (frames): VerbNet class frames in prag could be stored as one maximal structure plus allowed reductions, so each matched frame records which slots were suppressed or instantiated.
• I2 (world): speech events could always carry addressee and content entities, left unfilled or inferred when the sentence omits them, so later reasoning can ask "told to whom?".
• I3 (tree): a graded transitivity score per verb could serve as a prior when the parser must choose between a complement-clause attachment and an object reading for speech verbs.
• I4 (frames): role labelling could fall back to hierarchy-based defaults when a verb lacks a VerbNet entry, keeping the decision explainable.
• I5 (learn): alternation counts collected from silver parses could rank verbs by how generic they are within a class, which helps pick fallback heads in the ontology.