Home Recent Research People Papers Contact

What is Computational Meaning Dynamics?

An introduction to the research

A semantic observation is not simply a point. It is one observation along a trajectory generated by an underlying semantic process. Computational Meaning Dynamics (CMD) studies those trajectories: how meaning changes through context, sequence, interaction, and time, and what the geometry and dynamics of those changes reveal about stability, transition, and mechanism. Meaning representations are measurements of an underlying semantic state. However, the representation is the instrument and should not automatically be equated with the phenomenon being measured.

CMD starts from a simple but consequential premise: a semantic observation is not merely a point. It is one observation along a trajectory generated by an underlying semantic process. That changes the research question from how similar are these two representations? to how is meaning moving, what kind of transition is occurring, and what mechanism may be producing it?

Beyond semantic drift

Two semantic basins connected by a bridge, showing tangential movement within one basin, radial excursion, and a trajectory crossing into a new semantic regime.
Semantic trajectories can move without leaving a local meaning basin, while an alternative path through a bridge region can produce a substantive regime transition. Bounded radial departure is different from continuing tangential reinterpretation and illustrates why distance from the starting state alone is insufficient to characterize semantic change.

Much semantic-change work concentrates on displacement: whether a word, concept, document, model response, or other semantic object moved from one representation to another. CMD treats displacement as only one observable. Two trajectories can move the same distance while behaving very differently: one may remain inside a stable semantic regime, another may reorganize its neighborhood, another may oscillate, and another may cross into a new basin of meaning.

The goal is therefore not simply to detect that meaning changed, but to characterize how it changed. CMD distinguishes translation from reorganization, excursion from transition, bounded departure from continuing reinterpretation, local churn from regime change, and organic evolution from change introduced by an intervention or processing pipeline.

A meaning trajectory in practice

Consider an evolving incident initially described as an industrial accident. Early reporting clusters around equipment failure, casualties, and emergency response. Later reports introduce evidence of deliberate interference, responsibility claims, and attribution, shifting the reporting frame toward sabotage. A static comparison can show that the later reporting differs from the earlier reporting. CMD asks more. Did the interpretation move gradually or abruptly? Did the surrounding semantic neighborhood reorganize? Was the change a temporary excursion or a transition into a new semantic regime? Did the trajectory begin changing before the new frame became explicit? Those are dynamical questions, not simply similarity questions.

Measurement field concepts
One evolving phenomenon, several ways to see it: position, neighborhood, trajectory, and process describe different properties of the same semantic path rather than separate metrics.

Measurement before interpretation

CMD treats embeddings, graphs, distributions, and related representations as measurement instruments. Movement in embedding space can be evidence of semantic change, but it can also reflect the properties of the instrument: model geometry, alignment, sampling, frequency, windowing, or representation choice. The scientific problem is therefore not only to measure movement, but to determine which observed properties are stable across representations, scales, and measurement choices.

This makes representation invariance and external validation central to the research. A useful semantic observable should survive reasonable changes in the instrument and should correspond to independently observable changes in meaning or behavior rather than merely to movement inside one embedding space.

What CMD is trying to discover

A central question is which properties of meaning trajectories remain stable across representations and scales, and which of those properties identify independently observable semantic transitions better than static semantic similarity. A related question is whether the dynamics of a semantic trajectory expose structure, transitions, or future behavior that cannot be recovered from semantic position or pairwise similarity alone.

These questions lead to a broader research area: defining semantic state and measurement validity; characterizing trajectories, basins, bridges, recurrence, persistence, and regime transitions; testing the same constructs across different domains; and determining how interventions such as grounding, retrieval, context changes, or model changes alter the path that meaning takes.

A continuous semantic trajectory passes through basin-like regions representing measurement, dynamics, empirical testing, and intervention
A semantic trajectory moves continuously through a series of basin-like structures, linking four parts of CMD research: measurement theory, trajectory dynamics, empirical testing, and intervention and control. The changing loops, transitions, and perturbation points represent the recurring states, regime changes, and path alterations used to study what semantic dynamics reveal beyond static similarity.

A theory of the evolving semantic substrate

Semantic Substrate Dynamics Theory (SSDT) provides a formal foundation for this work. It represents an analysis window as a semantic substrate St = (X, dt, Pt), coupling a set of semantic objects, an embedding-induced geometry, and a local diffusion process. Within that common substrate, translational drift, neighborhood rewiring, recursive trajectory behavior, and intervention effects become different observables of the same evolving system rather than unrelated quantities all labeled as drift.

Geometry supplies a mechanism layer. Coarse Ricci curvature characterizes locally contractive, basin-like structure and expansive, bridge-like structure; trajectory decomposition separates bounded radial departure from continuing tangential reinterpretation; and operator composition makes it possible to study whether the timing and ordering of interventions change semantic outcomes. These constructs are used as testable descriptions and hypotheses, not as substitutes for empirical validation.

Why these dynamics matter

Systems increasingly act on representations whose meaning can change while the representation itself remains in use. An operational system can preserve the same event while its interpretation changes. A clinical code can remain syntactically identical while its meaning in use shifts. Repeated AI transformation can move an answer into a different semantic regime even though each individual step appears reasonable. Static similarity can reveal that two states differ. It cannot reliably tell whether the difference is harmless variation, continuing reinterpretation, an unstable excursion, or a transition that changes what the representation now means. Those distinctions affect when systems should be trusted, examined, grounded, updated, or allowed to continue operating without intervention.

CMD is an attempt to make those distinctions measurable. The goal is a computational science of changing meaning in which trajectories can be observed, transition mechanisms tested, measurements compared across representations, and the resulting insights turned into working tools.

CMD Lineage Our Research Areas Read the SSDT paper