Scalable and interpretable quantum natural language processing: an implementation on trapped ions

TL;DR

This paper demonstrates the first scalable, interpretable quantum natural language processing implementation on trapped ions, showcasing compositional generalisation and interpretability in a quantum question-answering model.

Contribution

It introduces a novel quantum NLP model with compositional structure, enabling interpretability and generalisation, and provides the first experimental proof on a trapped-ion quantum processor.

Findings

Successful implementation on Quantinuum's H1-1 quantum processor

Model exhibits compositional generalisation not achieved by classical models

Provides resource estimates for classical simulation of the quantum model

Abstract

We present the first implementation of text-level quantum natural language processing, a field where quantum computing and AI have found a fruitful intersection. We focus on the QDisCoCirc model, which is underpinned by a compositional approach to rendering AI interpretable: the behaviour of the whole can be understood in terms of the behaviour of parts, and the way they are put together. Interpretability is crucial for understanding the unwanted behaviours of AI. By leveraging the compositional structure in the model's architecture, we introduce a novel setup which enables 'compositional generalisation': we classically train components which are then composed to generate larger test instances, the evaluation of which asymptotically requires a quantum computer. Another key advantage of our approach is that it bypasses the trainability challenges arising in quantum machine learning. The main task that we consider is the model-native task of question-answering, and we handcraft toy scale data that serves as a proving ground. We demonstrate an experiment on Quantinuum's H1-1 trapped-ion quantum processor, which constitutes the first proof of concept implementation of scalable compositional QNLP. We also provide resource estimates for classically simulating the model. The compositional structure allows us to inspect and interpret the word embeddings the model learns for each word, as well as the way in which they interact. This improves our understanding of how it tackles the question-answering task. As an initial comparison with classical baselines, we considered transformer and LSTM models, as well as GPT-4, none of which succeeded at compositional generalisation.