Parameter-Efficient Quantum NLP for Paraphrase Detection: Performance, Robustness, and Entanglement

Farha Nausheen; Khandakar Ahmed; Farina Riaz

quant-phSep 13, 2026

Parameter-Efficient Quantum NLP for Paraphrase Detection: Performance, Robustness, and Entanglement

Authors: Farha Nausheen, Khandakar Ahmed, Farina Riaz

Organizations: ISILC, Victoria University, Melbourne, VIC, Australia · CSIRO, Australian International Institute of Higher Education, Sydney, NSW, Australia

Abstract

Rigorous empirical validation of quantum machine learning on natural language tasks remains scarce. We evaluate a 10-qubit hybrid quantum-classical variational circuit (2,148 parameters) for paraphrase detection across three benchmarks: MRPC, Quora Question Pairs (QQP), and adversarial PAWS. On QQP (n = 10 seeds), the circuit achieves 75.53% +/- 0.75%. accuracy, statistically outperforming parameter-matched classical baselines (DeepMLP: p = 0.015, Cohen's d = 1.20; F1: p < 0.001, d = 2.32) and surpassing DistilBERT-4bit with 31,191 fewer parameters. On MRPC the optimal 2-layer variant reaches 92% of BERT-base accuracy at 54,000 lower parameter cost. Circuit depth analysis reveals a dataset-depth scaling effect; entanglement analysis via the Meyer-Wallach measure identifies multi-qubit entanglement as the primary performance driver (r = 0.85 across four variants). Adversarial evaluation on PAWS reveals emergent robustness: 98.2% recall versus 81.6% classical (+16.6 pp, d = 1.24, p < 0.001), without adversarial training. These results constitute the first systematic parameter-matched multi-benchmark empirical validation of hybrid variational circuits for NLP. All results are from classical simulation; hardware validation is future work.

Figures & tables

Model	Depth	Params	MRPC Acc	MRPC F1	QQP Acc	QQP F1
Quantum v1	2L	2,028	75.98	83.28	69.20	13.48 ‡
Quantum v2	6L	2,148	67.94	76.46	75.53 $\pm$ 0.75	69.88 $\pm$ 0.66
Classical MLP	—	2,106	67.65	77.00	73.30	64.73
DistilBERT-4bit	—	66,000,000	68.40	$\approx$ 72	68.40	—
BERT-base (fine-tuned)	—	109,483,778	82.61	86.50	—	—

Table 1: MRPC and QQP-10K performance by circuit depth. QQP results for Quantum v2 are reported as mean

\pm

SD over 10 seeds

Figure 1: QQP-10K accuracy comparison ( $n=10$ seeds). Bar chart showing mean accuracy $\pm$ 1 SD for the quantum model (teal) and six parameter-matched classical baselines (grey). Quantum v2 achieves 75.53% — highest among all sub-3K parameter models. Welch $p=0.015$ , Cohen’s $d=1.20$ vs DeepMLP (accuracy); $p<0.001$ , $d=2.32$ on F1.

Model	Params	Accuracy (%)	F1 (%)	$\sigma$ acc
Quantum v2 (6L) $\star$	2,148	$\mathbf{75.53\pm 0.75}$	$\mathbf{69.88\pm 0.66}$	0.75
Quantum HEA	2,008	75.70 †	69.59 †	—
Quantum Amplitude	2,028	74.50 †	68.86 †	—
Quantum v1 (2L)	2,028	69.20 †	13.48 †‡	—
Classical DeepMLP	1,719	$74.54\pm 0.89$	$66.84\pm 1.73$	0.89
Classical Ensemble	1,998	$75.02\pm 2.09$	$66.48\pm 2.54$	2.09

Table 2: QQP-10K performance: all models (

n=10

seeds unless noted).

Figure 2: Cohen’s $d$ effect size: Quantum v2 vs. all classical baselines (QQP-10K, $n=10$ seeds). Top: horizontal bar chart of Cohen’s $d$ for each pairwise comparison; dashed vertical lines mark the small ( $d=0.2$ ), medium ( $d=0.5$ ), large ( $d=0.8$ ) and very-large ( $d=2.0$ ) thresholds. All comparisons exceed the large threshold; Quantum vs. CNN-1D reaches $d=4.50$ . Bottom: placement of each comparison on the Cohen’s $d$ magnitude interpretation scale (negligible $\to$ very large). The quantum model occupies the large-to-very-large region for every baseline, indicating robustly practical significance beyond mere statistical significance.

Figure 3: 95% confidence intervals for QQP-10K accuracy ( $n=10$ seeds). Diamond markers show mean accuracy; horizontal bars span the 95% CI (mean $\pm$ 1.96 $\sigma/\!\sqrt{n}$ ). The dashed line marks the quantum lower bound (75.07%), which lies above the upper bound of every classical baseline, confirming non-overlapping intervals.

Figure 4: Pairwise statistical significance heatmaps. Left: $-\log_{10}(p)$ values from two-sample Welch $t$ -tests; darker red indicates higher significance. Right: Significance level after Bonferroni correction ( $***$ $p<0.001$ , $**$ $p<0.01$ , $*$ $p<0.05$ , n.s. not significant). The quantum model (teal border) achieves $***$ against all six classical baselines.

Dataset	$N$	Depth	Q Acc	Q F1	C Acc	C F1	$\Delta$ Acc	$\Delta$ F1
MRPC	3,668	2L	75.98	83.28	67.65	77.00	$\mathbf{+8.33}$ pp	$\mathbf{+6.28}$ pp
MRPC	3,668	6L	67.94	76.46	67.65	77.00	$+0.29$ pp	$-0.54$ pp
QQP-10K	10,000	2L	69.20	13.48 ‡	73.30	64.73	$-4.10$ pp	$-51.25$ pp
QQP-10K	10,000	6L	75.53	69.88	74.54	66.84	$\mathbf{+0.99}$ pp	$\mathbf{+3.04}$ pp

Table 3: Circuit depth effect across datasets.

Figure 5: Dataset–depth scaling effect. Accuracy as a function of circuit depth (2L, 4L, 6L) on MRPC (3,668 samples, blue) and QQP-10K (10,000 samples, green). The curves cross between depths 2 and 6: the 2-layer circuit peaks on MRPC ( $+8.33$ pp over classical); the 6-layer circuit peaks on QQP-10K ( $+0.99$ pp). Shaded regions denote underfitting (left) and overfitting/barren-plateau risk (right).

Figure 6: Parameter efficiency: accuracy vs. model size. Log-scale scatter of parameter count vs. accuracy. The quantum model (teal diamond, 2,148 params, 75.53%) lies on the efficiency frontier above all classical baselines of comparable size and within ${\sim}1$ pp of BERT (109M params). The dashed line traces the efficiency frontier.

Circuit	Layers	MW $Q$	QQP Acc (%)
Quantum v1	2	$0.441\pm 0.031$	69.20
Quantum Amplitude	2	$0.523\pm 0.028$	74.50
Quantum HEA	3	$0.612\pm 0.025$	75.70
Quantum v2	6	$\mathbf{0.689\pm 0.023}$	75.53

Table 4: Entanglement vsȧccuracy across circuit variants.

Model	Acc (%)	Prec (%)	Recall (%)	F1 (%)	Ansatz / Topology
Quantum v2 (6L)	45.40	45.16	98.23	61.87	AngleEmb + all-to-all
Quantum HEA (2L) †	—	—	98.00	—	AngleEmb + linear chain
Quantum Amplitude (2L) †	—	—	98.00	—	AmplitudeEmb + circular
Quantum v1 (2L)	Not evaluated on PAWS				AngleEmb + circular
Classical MLP †	49.60	46.64	81.60	59.35	—
Classical Attention †	48.20	46.10	82.30	59.41	—

Table 5: PAWS adversarial robustness: all evaluated models (

n=1{,}000

adversarial pairs). Recall is the primary metric; bold indicates best per column.

\dagger

single-run result.

Figure 7: Adversarial robustness on PAWS ( $n=1{,}000$ ): PAWS recall by model. Horizontal bars show recall on PAWS word-scrambling adversarial pairs. All three evaluated quantum variants (purple: Quantum v2 6L all-to-all; cyan: Quantum HEA 2L linear; teal: Quantum Amplitude 2L circular) achieve $\geq\!98\%$ recall, independent of ansatz or connectivity. The classical MLP baseline achieves 81.60% (grey). The dashed vertical line marks the classical baseline; the red annotation indicates the $+16.6$ pp gap (Cohen’s $d=1.24$ , large effect). Quantum v1 (2L, circular) was not evaluated on PAWS. Statistical footer: $t=8.91$ , $p<0.001$ ; 95% CI quantum $[97.1\%,\,99.4\%]$ , classical $[79.2\%,\,84.1\%]$ (non-overlapping).

Panel of Figure 8 — Figure 8: Hybrid quantum-classical architecture. Top: four-stage pipeline. (1) Encoding : frozen SBERT ( all-MiniLM-L6-v2 ; 384 dim $\to$ 40 dim via PCA). (2) Classical pre-processing : FC(40 $\to$ 32) + BN + ReLU + FC(32 $\to$ 10). (3) Quantum circuit : 10-qubit AngleEmbedding + $L$ $\times$ StronglyEntanglingLayers ( $L=6$ for v2); all-to-all CNOT; PauliZ measurement. (4) Classical post-processing : FC(10 $\to$ 16) + BN + ReLU + Dropout(0.3) + FC(16 $\to$ 2). Total: 2,148 trainable parameters. Bottom: quantum circuit rendered directly from the PennyLane circuit definition via qml.draw_mpl() ; one representative variational layer shown ( $L=6$ total). Gates: RY (AngleEmbedding) followed by Rot $(\phi,\theta,\omega)$ + all-to-all CNOT(StronglyEntanglingLayers).

Component	Parameters
Pre-processing ( $W_{1}$ , $W_{2}$ + BN)	1,408
Quantum circuit ( $L=6$ layers)	$6\times 10\times 3=180$
Post-processing ( $W_{3}$ , $W_{4}$ + BN)	560
Total	2,148

Table 6: Parameter breakdown of the Quantum v2 model

Hyperparameter	Value
Optimiser	Adam
Learning rate	0.001
Batch size	8 (MRPC) / 32 (QQP-10K)
Max epochs	50
Early stopping patience	10
Dropout rate	0.30

Table 7: Hyperparameter configuration.

Figure S1: Seed-level stability analysis for Quantum v2 on QQP-10K ( $n=10$ seeds). Top-left: trial-by-trial accuracy (T1–T10) with mean (dashed) and 95% CI shaded band ( $[75.07\%,\,75.99\%]$ ); all 10 seeds lie within the band with no outliers. Top-right: empirical accuracy distribution with fitted normal ( $\mu=75.55\%$ , $\sigma=0.41\%$ ); Shapiro–Wilk $p=0.38$ (not significant), confirming the Gaussian assumption underlying Welch’s $t$ -test. Bottom-left: notched box plots for accuracy, F1, precision and recall ( $\sigma_{\mathrm{acc}}=0.39\%$ , $\sigma_{\mathrm{F1}}=0.37\%$ ), demonstrating consistent multi-metric stability. Bottom-right: temporal stability trace; coefficient of variation $=0.54\%$ , ruling out seed-ordering effects.

Figure S2: Multi-metric performance profile on QQP-10K (balanced evaluation). Radar chart comparing Quantum v2 (teal), DeepMLP (orange), and Classical MLP (dark blue) across Accuracy, Recall, Precision, and F1 Score. The quantum polygon encloses a strictly larger area than both classical baselines on every axis, with the most pronounced advantage on F1 (

+3.04

pp over DeepMLP and the smallest on Precision, consistent with the high-recall operating point on PAWS. Classical models overlap closely, confirming the quantum-classical gap is not explained by classical architectural differences. Quantum values are 10-seed means; classical values are 10-seed means where available.

Explore similar work

Hybrid Quantum-Classical NLP Classification with Compact Semantic Representations: An Experimental Analysis of Representation Compression
Sep 9, 2026Ali Hassan, Zijia Zhao, Maha A. MetaweiHybrid Quantum-Classical ML Text Classification
German University in Cairo, Cairo, Egypt · University of Melbourne, Melbourne, Australia · High Performance Computing Lab, Electronics Research Institute, Cairo, Egypt
SPSA Hyperparameter Tuning for Variational Quantum Natural Language Inference
Aug 11, 2026Nayan D'Souza, Christopher J. AgostinoHyperparameter Optimization Variational Quantum Circuits
NPC Worldwide, Bloomington, IN, USA · Department of Linguistics, Indiana University, Bloomington, IN, USA
Parameter-efficient Quantum Multi-task Learning
Apr 15, 2026Hevish Cowlessur, Chandra Thapa, Tansu Alpcan +1Multi-Task Learning Quantum Machine Learning
Department of Electrical and Electronic Engineering, University of Melbourne, Parkville, 3010, VIC, Australia. · CSIRO, Sydney, NSW, Australia.

Sep 9, 2026J/K move · Enter open · S save

Hybrid Quantum-Classical NLP Classification with Compact Semantic Representations: An Experimental Analysis of Representation Compression

Ali Hassan, Zijia Zhao, Maha A. Metawei

German University in Cairo, Cairo, Egypt · University of Melbourne, Melbourne, Australia · High Performance Computing Lab, Electronics Research Institute, Cairo, Egypt

Hybrid Quantum-Classical ML Text Classification Embedding Compression Dimensionality Reduction

Open paper AskPDF arXiv

Large language and sentence-embedding models provide rich semantic representations, but their high dimensionality poses a challenge for near-term quantum machine learning (QML), where quantum circuits can process only a limited number of input features. We investigate a hybrid quantum-classical pipeline that transforms high-dimensional sentence embeddings into compact representations for variational quantum classification. The workflow combines a pretrained sentence-embedding model, dimensionality reduction, angle encoding, a variational quantum circuit (VQC), and a classical decision layer. We systematically compare principal component analysis (PCA), neighborhood components analysis (NCA), and linear discriminant analysis (LDA), covering both unsupervised and supervised dimensionality reduction. Using the TREC question-classification dataset, we study the relationship between representation dimensionality, information retention, qubit count, and classification performance. Preliminary PCA experiments reveal a strong information bottleneck: reducing 768-dimensional embeddings to 3, 4, 5, and 8 dimensions retains about 8.2%, 10.2%, 11.9%, and 16.4% of the variance, with corresponding classification accuracies of 50.3%, 51.2%, 57.9%, and 63.4%. In contrast, supervised reduction is substantially more efficient. LDA reaches 85.3% accuracy and NCA reaches 83.1% using only 5 dimensions, under a leakage-free cross-validation protocol, compared with 85.1% for a full 384-dimensional classical baseline. These results indicate that supervised dimensionality reduction can preserve task-relevant information far more effectively than variance-based compression, making compact representations a promising route toward practical hybrid quantum-classical NLP models.

Abstract

Explore similar work

Hybrid Quantum-Classical NLP Classification with Compact Semantic Representations: An Experimental Analysis of Representation Compression

SPSA Hyperparameter Tuning for Variational Quantum Natural Language Inference

Parameter-efficient Quantum Multi-task Learning