Quantum Algorithms For Compositional Distributional Semantics in Natural Language Processing

Abstract

In this thesis, we try to combine machine learning and quantum computing in order to simultaneously perform different readings of a sentence. In the first part, we introduce a procedure to obtain a distributional tensor representation for adjectives seen as endomorphisms on the noun space. We accomplish this by discretizing the surface of the n-sphere, where the noun embeddings lie, and finding local approximations. We show how the analysis of their activity distributions can give us some information about their action in the noun space. We also provide different options to approximate their local behavior and visually represent their activity. This approach could, in principle, be applied to any kind of transformation in the sense of the types of categorial grammars. The second section, which is devoted to the quantum approach to the problem, explains how these classical representations can be effectively encoded into quantum states. Additionally, it involves the introduction of some quantum circuits, which, when used in conjunction with a traditional machine learning strategy, enable us to simultaneously perform several readings of ambiguous sentences using index contractions and to assign them a likelihood score.