Quantum exordium for natural language processing: A novel approach to sample on decoders

Abstract

The sampling task of Seq2Seq models in Natural Language Processing (NLP) is based on heuristics because of the Non-Deterministic Polynomial Time (NP) nature of this problem. The goal of this research is to develop a quantum sampler for Seq2Seq models, and give evidence that Quantum Annealing (QA) can guide the search space of these samplers. The contribution of this work is given by showing an architecture to represent Recurrent Neural Networks (RNN) in a quantum computer to finally develop a quantum sampler. The individual architectures (i.e. summation, multiplication, argmax, and activation functions) achieve optimal accuracies in both simulated and quantum environments. While the results of the overall proposal show that it can either outperform or match greedy approaches. As the very first steps of quantum NLP, these are tested against simple RNN with a synthetic data set of random numbers, and a real quantum computer is utilized. Since ane functions are the basis of most Artificial Intelligence (AI) models, this method can be applied to more complex architectures in the future.