Quantum Methods for Managing Ambiguity in Natural Language Processing

The fast-growing field of Quantum Natural Language Processing (QNLP: Widdows et al., 2024), in which the current article is situated, seeks to explain how information is processed in human language, using the mathematical framework of quantum theory. In QNLP, machine learning models are quantum circuits, which capture the meaning of sentences or other pieces of linguistic information. These models reflect an inherently compositional approach, in contrast to state-of-the-art machine learning models, such as deep neural networks, which renders them more interpretable (Coecke et al., 2020). The contributions of the current article are formulated in terms of diagrams in the Categorical Compositional Distributional (DisCoCat) framework (Coecke et al., 2010), in which word meanings are represented by tensors of various ranks. A noun, for example, is represented by a vector, whereas an intransitive verb is modelled as a matrix, and a transitive verb is represented by a rank-three tensor. The interaction of the meaning of these words, which results in the meaning of a sentence, is guided by the pregroup grammar. This combination of grammar and mathematical methods from tensor calculus allows the DisCoCat model to account for both the distributional and the compositional aspect of language. This connection is formally established via the mathematical framework of Category Theory . 1