Overview
This paper proposes that language is characterized by two fundamental parameters: amplitude and phase. Amplitude, broadly understood as the strength of association, is typically estimated by counting word co-occurrences, a mechanism refined by computational methods such as word embeddings and attention weights. These methods sum contributions from all writers within a corpus, effectively capturing a population-averaged state.
The core claim of this work is the existence of a second, distinct parameter: phase. Unlike amplitude, phase is conceptualized as existing exclusively between meanings. Its function is to determine the manner in which coactivated meanings combine, with the capacity to reverse the contribution of a meaning while that meaning remains fully present in the semantic landscape. Speakers are posited to set this phase within the linguistic signal through specific linguistic forms. Furthermore, individual encounters establish phase relations, and historical context subsequently distributes these relations across language use.
Research Context
The prevailing approaches in language modeling, particularly those utilizing traditional transformer architectures, are identified as lacking an explicit representation for phase during frozen inference. Current interpretability programs in this field, which measure progress through monosemanticity, are implicitly optimizing against the very phenomenon phase represents. The coexistence of multiple meanings, often treated as a defect in such models, is presented as the essential condition for phenomena like allusion, irony, and quotation. The paper argues that population averaging, inherent in agent-deindexed corpora, effectively deletes history-indexed phase information. Such corpora identify only the population marginal state, and consequently, cannot determine individual or dyadic states at any scale.
Approach
The research posits a theory that necessitates an explicit representation of phase. The claim defended is a 'weak version': that interpretation requires a second relational parameter, characterized as signed, persistent, and indexed to specific individuals and dyads. As a notational framework for this parameter, quantum probability is suggested, though the formalism does not claim the existence of quantum processes within the brain.
A 'strong version' of the claim is also articulated: that quantum calculus imposes constraints on these linguistic phenomena in ways that signed classical models cannot. This strong version, however, is contingent upon an encounter-order constraint that has not yet been derived.
Findings
The paper does not present empirical findings but rather outlines a theoretical framework and proposes six specific predictions designed to test its claims. These predictions address:
- Whether a suppressed meaning retains its activity.
- Whether the order of encounters influences the function of a phrase.
- Whether explicit marking within the signal alters how a shared phrase is interpreted.
- Whether a model provided with a history is fundamentally changed by it, or merely informed about it.
Why This Matters
The proposed architecture for future language models, according to this theory, would incorporate agent-indexed, phase-bearing semantic states. This suggests a redirection for computational linguistics and natural language processing, advocating for models that move beyond population-averaged statistics to account for individual and historical nuances in meaning construction. The distinction between amplitude and phase offers a conceptual framework for understanding how subtle linguistic cues contribute to complex interpretation, particularly in contexts involving non-literal or layered meanings.
Key Limitations Mentioned by Researchers
The 'strong version' of the claim, asserting that quantum calculus constrains these phenomena beyond the capabilities of signed classical models, is explicitly noted as resting on an encounter-order constraint that is not yet derived.