Quantum Contextuality for Contextual Word Embeddings

Abstract: Conventional word-to-vector embeddings face challenges in representing polysemy, where word meaning is context-dependent. While dynamic embeddings address this, we propose an alternative framework utilizing quantum contextuality. In this approach, words are encoded as single, static vectors within a Hilbert space. Language contexts are formalized as maximal observables, mathematically equivalent to orthonormal bases. A word vector acquires its specific semantic meaning based on the basis (context) it occupies, leveraging the quantum concept of intertwining contexts where a single vector can belong to multiple, mutually complementary bases. This method allows meaning to be constructed through orthogonality relationships inherent in the contextual structure, potentially offering a novel way to statically encode contextual semantics.