Pretrained Quantum-Inspired Deep Neural Network for Natural Language Processing.

Area of Science:

• Artificial Intelligence
• Quantum Computing
• Natural Language Processing

Background:

• Natural Language Processing (NLP) models often suffer from a "black-box" problem and fail to capture all linguistic features.
• Quantum-inspired models offer a potential solution but often neglect essential prior knowledge and pretrained text features.
• Developing interpretable and high-performing NLP models remains a significant challenge.

Purpose of the Study:

• To propose a novel pretrained quantum-inspired deep neural network for enhanced NLP.
• To address the limitations of existing quantum-inspired models by incorporating pretrained features.
• To improve both performance and interpretability in NLP tasks.

Main Methods:

• Developed a quantum-inspired pretrained feature embedding (QPFE) method to model word superposition states.
• Designed the QPFE-ERNIE model by integrating QPFE with semantic features from the ERNIE model.
• Evaluated the model on sentiment classification and word sense disambiguation (WSD) tasks, providing quantum circuit diagrams.

Main Results:

• QPFE-ERNIE significantly outperformed GRU, BiLSTM, and TextCNN in sentiment classification across five datasets.
• The model achieved superior accuracy, F1-score, and precision compared to ERNIE on CR and SST datasets.
• For WSD, QPFE-ERNIE showed notable improvements over BERT, ERNIE, and a previous quantum-inspired model (QWSD).

Conclusions:

• The proposed QPFE-ERNIE model offers a powerful and interpretable solution for NLP problems.
• This work lays the groundwork for future advancements in quantum-inspired NLP models.
• The integration of quantum theory with deep learning shows significant promise for NLP tasks.