Core Concepts
This paper surveys the state of quantum natural language processing, showing how NLP-related techniques including word embeddings, sequential models, attention, and grammatical parsing have been used in quantum language processing. It also introduces a new quantum design for the basic task of text encoding, and discusses the challenges of distinguishing between hypothetical and actual statements in language models.
Abstract
The paper begins by introducing some basic concepts in quantum computing, including superposition, entanglement, and quantum gates and circuits. It then presents a detailed example of how quantum circuits could be used to represent sequences of characters that make up natural language texts, highlighting both the promise and challenges of this approach.
The main body of the paper surveys various ways in which other aspects of language processing have been modeled on quantum computers, including word embeddings, sequential models, and attention mechanisms. For word embeddings, the paper discusses both memory-efficient and circuit-efficient approaches to encoding words as quantum states, and how these can be used in quantum versions of techniques like word2vec.
For sequential models, the paper reviews previous work on quantum versions of n-gram models and hidden Markov models, and proposes a new quantum recurrent neural network architecture that aims to balance expressivity and efficiency for current NISQ hardware.
The paper also discusses quantum approaches to attention mechanisms, including the Quantum Self-Attention Neural Network (QSANN) which seeks to leverage high-dimensional Hilbert spaces to extract correlations that are intractable classically.
Finally, the paper considers the challenges of distinguishing between hypothetical and actual statements in language models, noting that this problem has taken on fresh urgency in AI systems for fact-checking. It argues that quantum mechanics provides a better starting point than classical mechanics for modeling this distinction.