Quantum-inspired neural network with hierarchical entanglement embedding for matching

Area of Science:

• Computational linguistics
• Cognitive science
• Quantum computing applications

Background:

• Quantum-inspired neural networks (QNNs) show promise for modeling non-classical linguistic phenomena.
• Existing QNNs struggle with high memory/storage costs for complex embeddings.
• Current models do not fully exploit the word formation parallels with morpheme aggregation.

Purpose of the Study:

• To introduce a novel Quantum-inspired neural network with Hierarchical Entanglement Embedding (QHEE).
• To address memory and storage limitations in current QNNs for language understanding.
• To leverage finer-grained morphemes for efficient, multi-grained semantic representation.

Main Methods:

• Developed QHEE model utilizing intra-word and inter-word entanglement embeddings.
• Aggregated constituent morphemes from multiple perspectives using intra-word embeddings.
• Combined words via unitary transformation with inter-word embeddings to capture non-classical correlations.
• Employed morpheme embeddings, smaller in number and dimensionality than word embeddings, for parameter compression.

Main Results:

• QHEE demonstrated superior effectiveness compared to strong quantum-inspired baselines.
• The model achieved significant compression ability due to efficient morpheme-based embeddings.
• Experimental results validated the model's performance across four benchmark datasets and diverse downstream tasks.

Conclusions:

• QHEE effectively mitigates memory and storage issues in QNNs.
• The morpheme-based hierarchical entanglement approach enables efficient multi-grained semantic learning.
• QHEE represents a significant advancement in quantum-inspired language understanding models.