Scholarlink

We present QuCoWE, a framework that learns quantum-native word embeddings by training shallow, hardware-efficient parameterized quantum circuits (PQCs) with a contrastive skip-gram objective. Words are encoded by data-reuploading circuits with controlled ring entanglement; similarity is computed via quantum state fidelity and passed through a logit-fidelity head that aligns scores with the shifted-PMI scale of SGNS/Noise-Contrastive Estimation. To maintain trainability, we introduce an entanglement-budget regularizer based on single-qubit purity that mitigates barren plateaus. On Text8 and WikiText-2, QuCoWE attains competitive intrinsic (WordSim-353, SimLex-999) and extrinsic (SST-2, TREC-6) performance versus 50–100d classical baselines while using fewer learned parameters per token. All experiments are run in classical simulation; we analyze depolarizing/readout noise and include error-mitigation hooks (zero-noise extrapolation, randomized compiling) to facilitate hardware deployment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

QuCoWE: Quantum Contrastive Word Embeddings with Variational Circuits for Near-Term Quantum Devices

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rabimba Karanjai,
  • Hemanth Hegadehalli Madhavarao,
  • Lei Xu,
  • Weidong Shi

摘要

We present QuCoWE, a framework that learns quantum-native word embeddings by training shallow, hardware-efficient parameterized quantum circuits (PQCs) with a contrastive skip-gram objective. Words are encoded by data-reuploading circuits with controlled ring entanglement; similarity is computed via quantum state fidelity and passed through a logit-fidelity head that aligns scores with the shifted-PMI scale of SGNS/Noise-Contrastive Estimation. To maintain trainability, we introduce an entanglement-budget regularizer based on single-qubit purity that mitigates barren plateaus. On Text8 and WikiText-2, QuCoWE attains competitive intrinsic (WordSim-353, SimLex-999) and extrinsic (SST-2, TREC-6) performance versus 50–100d classical baselines while using fewer learned parameters per token. All experiments are run in classical simulation; we analyze depolarizing/readout noise and include error-mitigation hooks (zero-noise extrapolation, randomized compiling) to facilitate hardware deployment.