Vocabularies, Conversations, Contexts

Conversation Context (Paulo Vinluan)

[The book is available on Amazon and the O’Reilly digital platform.]

Related readings:

The swift and chaotic development of large language models and agentic systems has raised a number of issues regarding communication between autonomous agents. While technical issues can be addressed with the Model Context Protocol (MCP), issues arising from heterogeneous contexts remain subject to varying interpretations — some framed in terms of semantic layers, others in terms of knowledge graphs. Both understandings, however, can be framed in terms of meaning.

Language & Communication

Direct & Mediated Communication

The raison d’être of languages is to support communication, directly when vocabularies (labelled facts) can be unambiguously to mapped to meanings (concepts), or indirectly when communications are mediated through symbolic representations (categories).

This yields three semantic tiers: conversational, using thesauri to map vocabularies to meanings; organizational, using domains to map meanings to representations; and operational, using taxonomies to map vocabularies to representations. That’s assuming a single language.

Communication & Meanings

Beyond their sheer diversity, languages are often characterized by grammars combining four layers:

  • Lexical, for the meaning of individual terms
  • Syntactic, for the ways constructs (a.k.a. sentences) can be built and used
  • Semantic, for the meaning of constructs
  • Pragmatic, for the meaning of constructs in context

But there is a catch: with more than six thousand languages currently in use, there seems to be no limit to the intertwining of these layers — as if each language had picked its own idiosyncratic way to express meanings.

And yet, regardless of the sheer diversity of forms, languages are united in function: direct communication, when meanings can be grounded in bounded and specific vocabularies; mediated communication, which passes through open-ended symbolic representations. Historically, that divide maps onto the one between spoken and written language, with the latter having been developed to liberate the former from the constraints of immediate context — and, in doing so, to support the continuity and consistency of human organizations and endeavors. In other words: collective knowledge. The operational significance of this capability is illustrated by the postal service of the ancient Persian Empire, which transcribed messages between senders and receivers speaking different languages of the empire into a shared official Aramaic script. This scheme directly addresses the agentic semantic conundrum, which arises when conversations are set across heterogeneous operational and organizational contexts.

Vocabularies & Contexts

As exemplified by Persian epistolary exchanges, the meanings of a two-sided conversation are drawn from two kinds of sources: the vocabularies of the sender and receiver — Aramaic-speaking and Persian-speaking populations, for instance — and the statutory and socio-economic context of the Persian Empire. With ontologies, the meaning of conversations is addressed through vocabularies and contexts.

Controlled Vocabularies

Controlled vocabularies are obtained by combining domain-specific glossaries set within bounded grammatical contexts, assuming that meanings can be aligned at the concept level. Using SKOS, concepts are described using:

  • Preferred and alternative labels, possibly in different languages (‘Pipe’@en, ‘Pipa’@it)
  • Preferred and alternative symbols, possibly textual ones (e.g., “Pipe”)
  • Association (related), generalization (broader) or specialization (narrower) connectors
SKOS: Concepts can be associated with symbols (left) or literals (right)

That approach to vocabularies comes with caveats.

First, assuming a clear divide between syntax and semantics entails that meanings are defined at word level, possibly with regard to domains, but regardless of the roles of words in sentences.

Second, the absence of any divide between the lexical and semantic layers entails an isomorphism of meanings between concepts and words: there is no room for homonymous Pipe concepts, one for smoke and the other for fluids. The only way out is to make the SKOS lexical-semantic conundrum explicit — e.g., Pipe-fluid and Pipe-smoke.

These caveats point to the nature of the correspondence between the names of concepts and their essence — echoing Aristotle’s distinction between substance and accident.

The Meaning of Concepts

Concepts are human artifacts, represented differently depending on whether a language is logographic or alphabetic. And the differences run deeper than language, as concepts are also framed by societies and cultures. While SKOS allows for both logographic (symbols) and alphabetic (labels) representations, neither can be used to identify concepts on its own. Far from being a weakness, this caveat marks a productive divide: language, supported by SKOS, and knowledge, supported by knowledge graphs. Within ontological prisms, that level of indirection is managed by thesauri: meanings in vocabularies are organized at the lexical level (a), meanings in knowledge graphs are organized at the conceptual level (b), and shared semantics are uniquely defined in thesauri (c) by combining (a) and (b).

A primary benefit is enabling interoperability between vocabularies without forcing their alignment. As a corollary, thanks to thesauri gearing vocabularies to knowledge graphs, the nature of labelled facts and the meaning of concepts can change independently while their semantic consistency is preserved. That level of indirection paves the way to a principled definition of conversational contexts.

Making Contexts

As LLMs have become the jack-of-all-trades agentic communication appliance, exchanges with cognizant interlocutors tasked with knowledge-intensive work have taken center stage — and, consequently, so has the contexts supporting them. These contexts combine three remits: semantic, cognitive, and conversational.

  • The semantic remit secures bounded and unambiguous meanings of vocabularies (a).
  • The cognitive remit provides bounded and consistent knowledge graphs (b).
  • The conversational remit ensures a continuous and coherent recording of conversations (c).

Combined with ontological prisms’ knowledge architecture, that modular approach to contexts enables a smooth collaboration between agents sharing whole or parts of vocabularies or knowledge.

Using Contexts

When agents relying on ontological prisms collaborate they start walking through their respective vocabularies and try to align the terms considered (a?), and check meanings with their respective thesauri (b?).

When they are part of some professional or institutional environment they can add shared knowledge to the context (c?).

FURTHER READING

Kaleidoscope Series

Other Caminao References