The compositionality principle is a classical problem in semantics: “is the meaning of a
structure entirely determined by the meaning of its constituents?”. However, the very
object of semantics is elusive: is it about symbols, things, world states, cognitive or neural
states within an individual, or emerging structures in culture or communities or agents?
Or all of that, e.g. as a result of interaction between agents and their environments? (cf.
the heterogeneous contributions on the compositionality principle, ranging from
embodied to symbolic views, collected in [
Reasonably, compositionality could only be studied after fixing what are the constituents of a structure. Let’s concentrate on natural language semantics: if we fix constituents to syntactic phrase-based grammars, and we assume a semantic theory for phrases, compositionality effects on phrase semantics can be evaluated on the basis of grammatical rules, as with the sentence The cigar-shaped asteroid Oumuamua and the hypothesis of Harvard: “It can be an alien spacecraft”. Multiple compositional phenomena appear: cigar-shaped, cigar-shaped asteroid, hypothesis of Harvard, alien spacecraft, cigar-shaped asteroid Oumuamua and the hypothesis of Harvard, the hypothesis of Harvard: “It can be an alien spacecraft’, etc.
Unfortunately, current tools for natural language processing do not give us complete accounts of compositionality even at a phrase level: noun-noun compounds have opaque semantic relations, adjectival modifiers follow unpredictable patterns, parataxis (as provided by conjunctions and punctuation) is semantically underspecified, metonymy (as in Harvard) or metaphor (as in cigar-shaped) require knowledge beyond the typical one associated with lexical constituents, etc.
Indeed, these problems have emerged very early in artificial intelligence, e.g. robust
parsing [
However, optimising a cost function is not necessarily, or not only, what scientific research is about. In the case of semantic compositionality, science may want to address what are the possible constituents of semantics proper, and if compositionality relations correspond to what we make sense of when interpreting e.g. a linguistic construction.
Given such premises, the contribution of knowledge graphs, and especially linked data and ontologies, is to make it explicit, and eventually test, hypotheses about what constitutes semantics in a certain context, and link it to public identities of individuals, concepts, relations, models, and things in general. Once represented into knowledge graphs, candidate semantic constructions become the object of investigation, and patterns may emerge, and can be associated with other constructions represented under different constraints. For example, multiple news describing a same event, or a series of related events, may be reconciled, and the dynamics of events, their storytelling, judgments, participants, can be investigated on a rigorous basis.
FRED [
In the knowledge graph perspective, compositionality becomes a matter of model reconciliation, evolution, and its related dynamics. However, current ontologies and data hardly address higher levels of meaning, e.g. multi-modality, grounding, attitudes, metaphor, sentiment, emotion, normative principles, storytelling, interpersonal issues, etc.
A semantic theory that provides a clean, simple solution in order to talk about
different meaning levels is frame semantics [
The compositionality arising from the frames evoked in a a-priori way in
linguistic resources, or detected in large corpora, has been started to be investigated by using
hybrid symbolic and statistical tools [