📄 Representation of English Phrasal Verbs in Ontological Semantics

Summary

Phrasal verbs are hard for semantic lexicons for two reasons. Their particles can be separated from the verb in the sentence, and their meanings often cannot be predicted from the parts. This paper describes how English phrasal verbs (PVs) were added as a new "microtheory" to the Knowledge Base Acquisition Editor (KBAE), the working implementation of Ontological Semantics (OS).

The starting list was high-frequency PVs taken from the Longman Pocket Phrasal Verbs Dictionary and then checked against FrameNet. The author sorts verb-particle combinations into four groups: purely prepositional, compositional, metaphorically extended compositional, and non-compositional. Two OS-specific parameters separate true PVs from prepositional verbs: aspect and modality. Each entry links syntactic templates covering both word orders to an ontological event with case roles and constraints.

The result is a methodology and template set, not an evaluated system. No quantitative testing is reported. The author's main open problem is metaphorically extended meanings. The work matters as the first attempt to enter non-literal multi-word expressions into the OS ontology.

Key points

• KBAE has four static resources. These are the ontology (rooted in ALL, which splits into EVENTS, OBJECTS and PROPERTIES), the Fact Database (instance memory), the onomasticon (proper names) and the lexicon.
• Each lexicon entry for a PV has three zones:
  • anno: definition, example and comments.
  • syn-struc: the possible orders, with $varX variables.
  • sem-struc: the ontological concept, its roles and selectional limits, written as SEM and RELAXABLE-TO facets.
• A combination counts as a PV rather than a prepositional verb if it has at least one of four features:
  • an aspect feature, defined by phase and iteration;
  • a modality feature, with seven types, each valued from 0 to 1.0;
  • partly figurative meaning;
  • non-compositional meaning.
• Examples:
  • make up maps to CREATIVE-COGNITIVE-EVENT, with a LIE theme and evaluative modality below 0.9.
  • pick off maps to KILL, with phase begin-continue-end and multiple iteration.
• Template design:
  • Four syntactic templates were first proposed. One was dropped because its optional NP made representations ambiguous.
  • Three remain: V NP P / V P NP, V P NP, and V P.
  • "Heavy" PVs such as make away with get extra lines inside an existing template rather than a new template.

Interesting ideas

• I1. A single lexical entry lists both the joined and the split order of a separable PV, and shared variables tie each syntactic slot to the same semantic role, so word order never changes the meaning.
• I2. A particle's contribution can be stored as aspect values (phase and number of repetitions) instead of as a separate sense, as with pick off meaning repeated killing.
• I3. Selectional restrictions come in two layers: a preferred SEM constraint and a RELAXABLE-TO limit on how far it may be violated, which allows figurative uses without dropping constraints altogether.
• I4. A graded evaluative modality is attached to the event itself; for example, make up stories carries a built-in negative judgement.
• I5. Templates that introduce optional slots were dropped because they caused ambiguity, and rare complex cases are handled by extending existing templates rather than adding new ones.

For Mova

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

• I1 (frames): A lexicon entry keyed on verb lemma plus compound:prt particle could map either order to one frame, with roles bound through shared variables.
• I2 (world): Particles such as off, up or on could add iteration and phase marks to events, which would improve event counting and event ordering in the world of the text.
• I3 (frames): Role assignment from VerbNet could score a strict match first, then a relaxed one, and record which level was used, so metaphorical readings stay explainable.
• I4 (frames): Storing attitude values on verb frames would give the implicature and irony modules a lexical source for the speaker's stance.
• I5 (world): This keeps the compiled verb-class inventory small and unambiguous, while still covering multi-particle verbs like look up to.