Quantum Natural Language Processing (QNLP)

• Quantum Natural Language Processing (QNLP) is a field that combines quantum computing and linguistic analysis using categorical compositional semantics to enable quantum-native language processing.
• It translates grammatical and discourse structures into quantum circuits through diagrammatic methods such as string diagrams and Frobenius spiders, bridging formal semantics with quantum hardware.
• Empirical studies using pronoun resolution tasks demonstrate significant accuracy improvements, validating QNLP's potential to handle complex multi-sentence discourse phenomena.

Quantum Natural Language Processing (QNLP) is a field at the intersection of quantum information theory, categorical compositional semantics, and natural language processing. QNLP leverages quantum computational models—especially those inspired by the mathematical structure of quantum theory and categorical linguistics—to represent, process, and reason about linguistic data in ways that are natively compatible with quantum hardware. This approach promises quantum-native linguistic inference, including handling discourse-level phenomena such as anaphora and ellipsis, via explicitly diagrammatic and compositional models mapped efficiently to quantum circuits.

1. Logical and Semantic Foundations

Quantum Natural Language Processing in the cited work (Wazni et al., 2022) is grounded in the non-commutative Lambek calculus extended by two crucial modalities: the "soft" exponential ($!$), enabling bounded contraction, and a "storage" modality ($\nabla$), allowing permutation of components. The language of formulas is: $$A,B ::= p \mid A\cdot B \mid A\backslash B \mid B/A \mid !A \mid \nabla A$$ with atomic types $p\in\{n,s\}$ for noun-phrase and sentence.

Key Gentzen-style sequent rules include:

• Identity: $A \Rightarrow A$
• Product Left/Right: Decomposition and combination of types via $\cdot$
• Division Left/Right: Usual non-commutative residuation
• Soft exponential (!): $!_L$ restricts contraction to at most $k_0$ instances, enforcing decidability and reflecting soft linear logic.
• Storage ($\nabla$): $\nabla_L$ and $\nabla_R$ rules, along with permutation rules, enabling non-strict word ordering found in natural language.

Semantically, each atomic type is assigned a finite-dimensional vector space, e.g., $[n]=N$, $[s]=S$, with composite types constructed via tensor product and dualization: $$[A\cdot B] = [A]\otimes [B],\quad [B/A] = [B]\otimes [A]^*,\quad [A\backslash B] = [A]^*\otimes [B]$$ For the exponential modality, truncated Fock space semantics are invoked: $$[!A] = \mathcal{F}_{k_0}([A]) := \bigoplus_{i=0}^{k_0}[A]^{\otimes i}$$ ensuring storage of at most $k_0$ copies and facilitating learnability from corpora. Proofs correspond to linear maps in FdVect under the categorical semantics.

String diagrammatic notation is used to represent derivations, encoding ordinary vector spaces as thin wires, Fock spaces as bold wires labeled by type, projections as explicit boxes, and contractions/cups/caps according to Selinger's graphical calculus.

2. Quantum Circuit Mapping and Implementation

QNLP translates these categorical linguistic structures to quantum circuits via a DisCoCat-to-quantum pipeline. The concrete translation steps are:

• Quantum Encoding of Types: Each wire of dimension $d$ is mapped to at least $\lceil \log_2 d \rceil$ qubits; one qubit per atomic type suffices in the simplest implementation.
• Word-State Preparation: Each vector $v\in V$ is encoded as a parameterized quantum state $|v(\theta)\rangle$ constructed via single-qubit rotations and Hadamards, e.g., $|v(\theta)\rangle = R_z(\theta) H|0\rangle$.
• Fock Space Contraction: The $p_n$ projection extracts a specific tensor-power component from the Fock space, operationalized as passing to $n$-qubit tensor states within the circuit.
• Grammatical Structure and Entanglement: Tensor products correspond to parallel preparation of qubit registers. Grammatical contractions, e.g., via cups, induce entangling gates (typically CNOT or CZ), possibly eliminated or compiled out for circuit depth efficiency.
• Spiders and Combination Maps: Higher-order contractions such as Frobenius spiders are mapped to CNOT-based structures; an alternative is controlled-$R_z$ (“IQP ansatz”) for parameter-efficient entanglement.

This translation is practically supported by toolkits such as DisCoPy and lambeq, integrating all parsing, rewriting, ansatz mapping, and export to quantum circuits.

3. Discourse-Level Pronoun Resolution Task

The core empirical contribution is a quantum implementation of a pronoun resolution task inspired by Winograd schemas. The dataset comprises 144 discourses of the form

• S₁: "The X V₁ Y."
• S₂: "They/It V₂ A."

The pronoun in S₂ ambiguously refers to either subject (X) or object (Y), resulting in binary classification: subject-reference (class 0) or object-reference (class 1). The dataset is balanced (72 train, 36 validation, 36 test) with vocabulary expansion performed via BERT-masking.

Each discourse is encoded classically as a pair of string diagrams (from SLLM parses), translated into quantum circuits. Pronoun-to-antecedent links are explicitly represented by parameter-tying and Fock-space projections. Two types of sentence-merging circuits are explored:

1. Frobenius (spider): CNOT-based composition, representing the diagrammatic "dot" fusion.
2. IQP ansatz: Controlled-$R_z$ between sentence outputs.

The full SLLM model ties pronoun and antecedent wires and parameters, aligning the quantum semantics of the resolved anaphora directly with the underlying discourse structure.

4. Model Training, Evaluation, and Results

Optimization is performed over the set of all word and circuit parameters $\Theta$ using the SPSA algorithm, suitable for quantum hardware or noisy simulators. For each circuit representing a (S₁, S₂) pair with label $y \in \{0,1\}$, the quantum measurement yields probabilities

$$p_i = |\langle i|\psi(\Theta)\rangle|^2 + \epsilon,\quad \epsilon = 10^{-9}$$

forming a two-component normalized vector $p=(p_0,p_1)/\sum p_i$, compared to one-hot $y$ via binary cross-entropy.

All models are simulated with 1,024 shots per evaluation on the IBMQ AerSimulator. Eight experimental conditions are evaluated:

• M1: No grammar, no discourse
• M2: No grammar, with discourse (pronoun link)
• M3: Grammar, no discourse
• M4: Grammar with discourse

Each is run with both spider and IQP-Rz combiners.

Best test set accuracies (averaged over 20 random runs):

Model	Spider (M⊙)	IQP‐Rz (M Rz)
1 (no-g,no-d)	64.44%	51.52%
2 (no-g,disc)	100.00%	67.70%
3 (g,no-d)	72.84%	48.97%
4 (g,disc)	91.38%	76.67%

Resolving discourse-level anaphora yields the largest improvement (up to +35%), particularly in spider models. Incorporating grammar adds moderate benefit, and Frobenius spiders consistently outperform IQP-controlled $R_z$ combiners. Training curves demonstrate stable convergence in all spider models, while IQP models occasionally stall around chance.

5. Diagrammatic Analysis and Quantum Insights

The use of string diagrams and explicit Fock-space projections in the quantum embedding allows for white-box interpretation of discourse links and circuit composition. Wires corresponding to stored antecedents are projected, permuted, and directly linked to pronoun qubits, matching formal semantics from SLLM. Parameter tying along these links simulates the effect of anaphoric binding at quantum-circuit level.

This approach shows that categorical quantum models are not limited to single-sentence semantics: the Fock-space plus spider/entangler extension enables explicit modeling of multi-sentence discourse phenomena, such as pronoun reference and possibly ellipsis.

Resource analysis indicates scaling to longer or more complex discourses will require additional qubits and increased circuit depth, with noise mitigation becoming essential for real NISQ devices.

6. Broader Implications and Future Directions

This work establishes three advances for QNLP:

1. Extension of SLLM (soft linear logic with modalities) to quantum-circuit semantics, encompassing discourse-level structures.
2. Demonstration of full end-to-end quantum simulation (from proof theory to classification) for pronoun resolution.
3. Empirical validation of discourse structures and Frobenius spiders as effective quantum-combinatorial primitives for linguistic compositionality.

Future work targets larger and more complex datasets, experiments on physical quantum hardware, tackling ellipsis via discourse circuits, and hybrid architectures fusing classical pre-training with compositional quantum layers.

A broader implication is the systematic, diagrammatic, and compositional approach promoted in QNLP: this enables both scalability (via compositional generalization) and explicit interpretability, separating QNLP from parameter-heavy and empirically opaque black-box methods. The result is a rigorous path toward quantum-native natural language understanding with full compositional transparency (Wazni et al., 2022).