Text classification using quantum Word2Vec vectorization

Abstract

This article is dedicated to the formation of quantum features for textual data, with a proposed LSTM-based approach for sarcasm detection. In the proposed method, words are first encoded in a quantum feature space, where semantic similarity is computed through quantum interference. Based on the encoded representations, quantum Word2Vec vectors are constructed and fed into a neural network to perform binary classification. The obtained results were compared with those of the classical approach. In the experiments, the quantum-based model achieved an accuracy of 97%, while the classical model reached 92%. Furthermore, the quantum approach required 309 seconds for training and 1 second for inference, whereas the classical approach required 412 seconds and 5 seconds, respectively. Thus, the computational time was reduced by approximately 25%–80%. The results demonstrate that the quantum approach, owing to its superposition and parallelism properties, achieves superior performance. All experiments were conducted using computer-based quantum simulators. Nevertheless, quantum Word2Vec and similar high-accuracy techniques show strong potential in natural language processing tasks, particularly in sarcasm detection.