Abstract
This article is an exploratory account of the non-monotonic behaviour of conceptual associations in the light of context. Computational approximations of conceptual space are furnished by semantic space models which are emerging from the fields of cognition and computational linguistics. Semantic space models not only provide a cognitively motivated basis to underpin human practical reasoning, but from a mathematical perspective, they are real-valued Hilbert spaces. This introduces the highly speculative prospect of formalizing aspects of human practical reasoning via quantum mechanics. This account focuses on how to formalize context effects in relation to concepts as well as keeping an eye on operational issues.
Introduction
Traditional Artificial Intelligence has long relied on symbolic logic to model human reasoning. However, "practical reasoning"—the way humans navigate everyday concepts and associations—often exhibits non-monotonic behavior that is difficult to capture with rigid rules. This paper investigates a novel intersection: using the mathematical framework of quantum mechanics to model how context dynamically shifts conceptual meanings within a high-dimensional "semantic space."
The authors propose that human reasoning is better modeled as a geometric process. By treating words as quantum states in a Hilbert space, they demonstrate how context acts like a "measurement" that causes a word's meaning to collapse from a mixture of possibilities into a specific sense.
The Problem: Beyond Symbolic Logic
Artificial Intelligence utilizes Non-Monotonic Reasoning (NMR) to handle situations where new information retracts previous conclusions (e.g., learning that a bird is a penguin and thus cannot fly). While theoretically sound, symbolic NMR struggles to scale to the complexity of real-world language. The authors argue that symbolic logic is merely the "tip of the iceberg." Beneath it lies a conceptual level where information is represented geometrically. The challenge is modeling how context (e.g., seeing "Iran" near "Reagan") dynamically shifts associations in a way that is both computationally tractable and cognitively plausible.
Methodology: Constructing Semantic Space
The paper bridges cognitive science and quantum mechanics using several key computational tools:
- Hyperspace Analogue to Language (HAL): This method builds semantic spaces by scanning large text bodies with a sliding window to record word co-occurrences. This creates a high-dimensional matrix where each word is represented as a vector.
- Hilbert Spaces and Dirac Notation: The resulting semantic spaces are treated as real-valued Hilbert spaces. Words are represented using Dirac notation as "kets" ($|w\rangle$).
- Singular Value Decomposition (SVD): This technique decomposes a word’s association matrix into its eigenstates, which represent the different "senses" or potential meanings of a word.
- Quantum Collapse: The core innovation is modeling context as a measurement operator. When a word $|v\rangle$ is encountered within the context of another word $u$, it undergoes a "collapse" into a new state $|v_u\rangle$, shifting its association weights.
Experimental Case Study: "President Reagan"
To test this model, the authors utilized the Reuters-21578 news collection from 1988. This dataset provided a "global" semantic space to analyze associations surrounding the word "Reagan."
- Words as Superpositions: In isolation, a word is a "mixture of potentialities." In the 1988 corpus, "Reagan" exists in a superposition of senses including "President," "Republican," and "Iran-Contra scandal."
- Context as Measurement: Seeing "Reagan" in the context of "Iran" is mathematically analogous to a quantum measurement. It forces the meaning to collapse. While the "global" Reagan is associated with "President" and "Administration," the "collapsed" Reagan in the context of "Iran" brings associations like "arms," "scandal," and "contra" to the forefront.
- Non-Symmetry: The model correctly predicts that context is asymmetric; the associations of "Reagan" given "Iran" differ from those of "Iran" given "Reagan."
Discussion and Future Work
The authors found that context often results in a "partial collapse." For instance, "Reagan" in the context of "Iran" still reflected a mixture of the "Iran-Contra" sense and the "Iran-Iraq War" sense, mirroring the complexity of human thought.
However, several challenges remain:
- Orthogonality: Unlike standard quantum mechanics, different linguistic senses often overlap rather than being perfectly independent.
- Entanglement: The authors suggest that complex concepts (like "pet fish") might require modeling via quantum entanglement, where the combination of concepts creates a state inseparable into its parts.
- Scalability: While the mathematical proof-of-concept is successful for individual examples, scaling this to a full "Quantum Logic" reasoning system remains a significant engineering hurdle.
In conclusion, this research suggests that the "logic" of human reasoning may not be a set of rules we follow, but a byproduct of the geometric structure of our conceptual space and the way context triggers state collapses.
Summary updated