Experimental Quantum Natural Language Processing for the Travel Industry

In the last decade, Natural Language Processing (NLP) has made giant leaps in treating textual data, by means of large deep neural networks capable of addressing complex tasks such as machine translation, language generation and text summa- rization. These models are based on the principle of distributionality — a word’s meaning is defined by the context in which it appears — and learn a represen- tation of words in a vector space. This implies training the model on a massive amount of data to learn the interdependencies among words. In addition, despite the successful achievements, the general problem of natural language understanding is still unsolved as neural network models lack of explainability, which raises further concerns when we entrust our decisions to their predictions in critical domains. Another line of research, sparked by linguists such as N. Chomsky and J. Lam- bek and culminated in the distributional compositional categorical (DisCoCat) model of Coecke, Clark and Sadradeh tries to address natural language under- standing from a different perspective, introducing a simple, yet powerful, analogy between computational linguistics and quantum theory: the grammatical structure of text and sentences connects the meaning of words in the same way that entan- glement connects the states of a quantum system. This language-to-qubit analogy is mathematically formalized using category theory. Starting from these foundations, this thesis work engineers the process of using quantum computers for natural language processing on real-world data related to the travel industry, assessing at the same time the expressiveness of the approach and the maturity of the technology it is destined to, providing experiments on and benchmarks of real world quantum hardware and showing that noisy intermediate- scale quantum (NISQ) devices are Quantum-NLP-friendly. Lastly, we provide com- parisons with popular state-of-the-art models and we show interesting advantages in terms of model size.