Part 4 of Foundations: events, roles, and why a sentence isn't a bag of triples.
When you sit down to store what a sentence says, a tempting representation is a set of bare subject-verb-object triples, the three-part records many knowledge graphs are built from. "Maria sold the bakery to Tom in 2023" might become Maria-sold-bakery. Where do Tom and 2023 attach? Extra triples can represent them, but without an identity for the selling event, repeated sales make those attachments ambiguous. This is a limitation of the bare encoding, not of graph or triple stores in general.
Linguistics offers a well-established alternative, by people who thought about sentences far longer than I have.
In "The Logical Form of Action Sentences", first published in 1967, Donald Davidson introduced an event variable to which participants and modifiers can attach. On that view there was a selling event: Maria was its agent, the bakery its theme, Tom its recipient, and 2023 its time. A second sale gets a second event identity and its own participants.
Charles Fillmore's case grammar and frame semantics, theories of how a verb frames the roles its participants play, added a complementary lens: predicates set expectations about their participants. A selling commonly has a seller, something sold, and a buyer, with optional details such as price and time. PropBank annotated predicate-argument roles across a corpus, while VerbNet organized verbs into classes with shared syntactic and semantic behavior. Resources such as AMR (Abstract Meaning Representation, a scheme for writing a sentence's meaning as a small graph) and precision grammars show different, complementary ways to make sentence structure explicit. None removes ambiguity from language; each gives an implementation a more disciplined place to represent it.
Keeping participants together#
Multi-participant questions become better specified. "Who did Maria sell the bakery to?" can constrain the seller, item, and recipient on the same event, rather than hoping an unattached "to Tom" fragment belongs to the right sale.
The whole disambiguates the parts. Knowing this is a selling frame makes a recipient reading of "to Tom" more plausible than a destination reading. The frame constrains the interpretation, although context can still override an expected role. This is the same lesson as part 2's questions, seen from the other side: wholes constrain parts.
Context also exists above the event. Words sit in clauses, clauses in sentences, and sentences in documents, and each layer can change what the ones below it mean. A pronoun may need the whole document to resolve; a relative date like "next quarter" may need a timeline. Keeping those links makes later review possible, even when interpretation is uncertain.
The resulting design rule is to avoid letting a claim fragment lose its context. Retain its connection to the event, sentence, document, and source span. That costs storage and implementation complexity, but it preserves information a later question or audit may need.
There is one more thing a sentence carries that no role label captures: how much force the speaker put behind it. On August 3 the series continues with the smallest words in the language, the mights and did-nots and cans, and what a memory owes them.