📄 Lexical decomposition of English prepositions and their fusion with verb lexical classes in motion constructions

Summary

The paper challenges a common Construction Grammar position, chiefly Goldberg's, that the path meaning in motion sentences comes from the construction itself and that the prepositions inside it add nothing. Working within the Lexical Constructional Model (LCM), the author proposes a tentative Lexical Template (LT) format and meta-language for English spatial particles. A semantic module applies Lexical Functions to a semantic primitive, adding topological, dynamic and functional detail to yield more specific items. A syntactic module records the situation type and the argument structure. Using COCA as the evidence base, the paper tests which particles combine with which Aktionsart verb classes in the Intransitive Motion and Caused Motion constructions. The main finding is that into, onto and over carry a change-of-location component themselves. That component lets telic or causative verbs that do not express motion, such as melt, shatter and scare, appear in motion constructions. Particles like at, on and under lack it, so they do not license these verbs. This matters because it locates part of the motion meaning in the lexicon rather than in the construction.

Key points

• The supplied text is the December 2011 article by Ignasi Navarro i Ferrando (Language Value, Vol. 3, No. 1, pp. 114–137), whose original title refers to how particles of space are absorbed into constructions expressing motion. This differs from the 2019 title and metadata given in the request.
• The LT format combines several frameworks:
  • Role and Reference Grammar logical structures
  • NSM-style primitives (BE, HAPPEN, BECOME, MOVE…)
  • Mel'čuk-style Lexical Functions, used paradigmatically (e.g., MAGN, INTENT, CONT, PURP)
  • Vendler/Van Valin Aktionsart classes
• The two constructional templates are:
  • Intransitive motion: do'(x) followed by BECOME be-LOC'(x, z)
  • Caused motion: do'(x, y) CAUSE BECOME be-LOC'(y, z)
• Corpus co-occurrence of each particle–verb-class pair is graded on a five-step scale: no, ?, yes?, yes, YES.
• Activity and active-accomplishment verbs (run, come, get, put) take most particles. With these verbs, in and on yield a motion reading only when it is ambiguous with a locative one.

Interesting ideas

• I1. Define a hyponym as its superordinate primitive plus a stack of named Lexical Functions (e.g., watch = see' with intensification, intent and prolonged continuity), which makes fine lexical distinctions explicit and composable.
• I2. A motion reading is licensed only if some word (the verb or the particle) supplies change of location, because the construction cannot coerce it alone.
• I3. Treat a spatial particle as a secondary predicate relating a trajector and a landmark, whose arguments may be shared with the main verb's predication.
• I4. Record graded corpus attestation per particle × Aktionsart cell, including "dubious" and "single instance" levels, instead of a binary allowed/forbidden judgement.
• I5. In motion contexts, in/on are ambiguous with locatives, whereas into/onto unambiguously mark an endpoint.

For Mova

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

• I1 (world): Verb-class knowledge could store each verb as a primitive plus function tags, so the ontology can explain how look, watch and observe differ rather than just listing them as near-synonyms.
• I2 (frames): When prag builds a motion frame, it could check for an explicit source of the "becomes located at" component and otherwise fall back to a locative frame, logging which word licensed the choice.
• I3 (world): The world model could add a separate location-state fact (trajector at or in the landmark) linked to the event, so "scared the kid onto the sidewalk" yields both a fear event and a new position.
• I4 (learn): A similar matrix built from silver data could serve as a soft prior and as negative controls, for example flagging parses that read "melt cheese on the croutes" as caused motion.
• I5 (back): During generation, prefer into or onto whenever the graph encodes a completed change of location, so the regenerated text does not drift into a static locative reading.