Technical Summary: Quantum Compositional NLP for Arabic

Problem Statement

Classical natural language processing (NLP) models, particularly those based on word embeddings and transformers, often treat sentences as "bags of words," discarding word order and relying on lexical co-occurrence statistics. While effective for languages with rigid syntax, this approximation fails severely for Arabic, a morphologically rich language with free word order (allowing Subject-Verb-Object, Verb-Subject-Object, and Nominal structures) and a complex root-and-pattern morphology system.

The paper posits that Arabic's structural complexity provides a unique testbed for Quantum Compositional Natural Language Processing (QNLP). Specifically, it investigates whether the DisCoCat (Discourse Compositional Category) framework, which maps grammatical structure to quantum circuit topology, can encode structural information in a way that is causally distinct from classical lexical models. The core challenge is to demonstrate that quantum circuits can distinguish between sentences with identical vocabulary but different grammatical structures, and that this distinction arises from the circuit's topology and entanglement rather than learned lexical parameters.

Methodology

Theoretical Framework

The system utilizes pregroup grammar, a type-theoretic formalism where words are assigned grammatical types (e.g., nouns as $n$, transitive verbs as $n^r \otimes s \otimes n^l$). A sentence is grammatical if the tensor product of its word types reduces to a sentence type $s$ via "cup" operations. In the quantum setting (DisCoCat), these string diagrams are compiled into quantum circuits:

• Words become qubit registers.
• Grammatical dependencies become entangling gates.
• Sentence meaning is a measurement outcome.

Crucially, the paper leverages the formal correspondence between Arabic's root-and-pattern morphology (where consonantal roots and vowel patterns operate on parallel information streams) and quantum tensor products (the formalization of parallel composition).

Pipeline Architecture

The authors developed a custom pipeline to connect Arabic to quantum circuits, addressing challenges such as right-to-left script, cliticization, and variable word order:

1. Morphological & Syntactic Analysis: Uses CAMeL Tools for morphological features (root, pattern, aspect) and Stanza for dependency parsing.
2. Structure Detection: Classifies sentences as SVO, VSO, or Nominal.
3. Diagram Construction: A custom module (arabic_dep_reader.py) generates pregroup diagrams. Notably, VSO sentences require a Swap operation in the diagram to align verb types ($s \otimes n^l \otimes n^l$) with their arguments, creating a topologically distinct circuit compared to SVO sentences ($n \otimes (n^r \otimes s \otimes n^l) \otimes n$).
4. Circuit Compilation: Diagrams are compiled into Instantaneous Quantum Polynomial (IQP) circuits using the lambeq library.
   • L0 (Depth 0): No entangling gates; word qubits evolve independently.
   • L1/L2: Parameterized controlled-Z rotation gates introduce entanglement.
5. Evaluation: Two strategies are employed:
   • Quantum Feature Map (QFM): Parameters are fixed to word embedding values; an SVM classifies the output.
   • SPSA (Simultaneous Perturbation Stochastic Approximation): Full end-to-end training of circuit parameters.

Experimental Design

Three controlled experiments were conducted to isolate specific structural properties:

1. Word Order (Experiment 1): A binary classification task (SVO vs. VSO) using 120 matched-pair sentences. Crucially, the same three words appear in both orders, ensuring that any model exceeding 50% accuracy must rely on structural information, not lexical identity.
2. Morphological Tense (Experiment 2): Binary classification of Past vs. Present tense verbs, where the signal is primarily lexical (different surface forms).
3. Word Sense Disambiguation (Experiment 3): A vocabulary-controlled dataset of 200 sentences involving four polysemous verbs. The design uses exact matched pairs and shared object/subject pools to isolate structural disambiguation signals from lexical ones.

Key Contributions

1. First Arabic QNLP System: The implementation of a pregroup grammar-based QNLP system for Arabic, including specific grammar rules for SVO, VSO, and Nominal structures, and a pipeline integrating Arabic morphological analyzers with quantum circuit compilers.
2. Causal Ablation of Entanglement: A controlled experiment demonstrating that parameterized entanglement is the sole cause of performance gains in structural tasks.
   • L0 Baseline: Circuits with grammar topology but no entanglement achieved exactly 50.0% accuracy on the matched-pair word order task with zero variance across all seeds and folds. This proves that without entanglement, the circuit cannot encode sentence-level structural differences, regardless of the topological difference in the diagram.
   • L1 Result: Adding a single layer of entangling gates raised accuracy to 64.9% (mean ± 3.2% std). The 15-percentage-point gain is causally attributed to entanglement.
3. Vocabulary-Controlled WSD Dataset: The creation of the first Arabic word sense disambiguation dataset using matched pairs and shared lexical pools to rigorously test structural vs. lexical signals.
4. Characterization of SPSA Label Inversion: The identification of a phenomenon where SPSA training converges to an inverted solution (correct separation, wrong orientation) in symmetric binary tasks. The paper demonstrates that ancilla qubit encoding (tracing out an ancilla to produce a density matrix) measurably reduces this inversion rate (e.g., from 99% to 23% on the qata'a verb task).

Results

• Word Order:

  • AraVec (Bag-of-Words): 12.8% (below chance due to spurious anti-correlations in matched pairs).
  • Topology-Only (0 params): 35.8% (raw), which implies 64.2% accuracy after correcting for an inverted decision boundary, confirming the topological signal exists but is hard to access without entanglement.
  • QFM L0: 50.0% (Zero variance; structural theorem).
  • QFM L1: 64.9% (CI [62.8, 66.3]).
  • AraBERT (Fine-tuned): 100% (Oracle upper bound, relying on positional encodings).
  • Learning Curves: As training data increased, QFM L1 performance improved (56% $\to$ 67%), while AraVec performance degraded (46% $\to$ 19%), confirming that the quantum model extracts structural signals while classical models fail on matched pairs.

• Morphological Tense:

  • Classical models (AraVec: 87%) outperformed quantum models (QFM: 56%, SPSA: 46.8%). This confirms that when the signal is lexical (different word forms), classical embeddings are superior, and the quantum circuit topology is less informative for this specific task.

• Word Sense Disambiguation:

  • Results were mixed. QFM generally performed near chance (47.4% pooled), as the structural signal for sense disambiguation is encoded in parameter values rather than topology.
  • SPSA achieved above-chance performance after symmetric correction (e.g., 79.5% for qata'a), demonstrating that training can align the circuit with subtle argument-structure features.
  • Ancilla encoding significantly reduced SPSA label inversion rates on symmetric tasks.

Significance and Claims

The paper claims that its primary significance lies not in outperforming state-of-the-art classical models (which AraBERT does easily), but in providing a causally identified, interpretable structural signal that is fundamentally different from classical mechanisms.

1. Mechanistic Distinction: The zero-variance L0 ablation proves that the quantum circuit's ability to distinguish word order is not an artifact of learning lexical statistics but is a direct consequence of entanglement acting on a grammar-derived topology.
2. Theoretical Justification for Arabic: The paper argues that Arabic is the ideal language for QNLP because its root-and-pattern morphology formally corresponds to parallel composition (multi-tape automata), which maps naturally to quantum tensor products. This structural alignment is unique among languages currently in the QNLP literature.
3. Low-Resource Potential: By encoding structural knowledge into circuit topology rather than relying on massive training corpora, QNLP offers a pathway for high-quality Arabic NLP in low-resource conditions.
4. Methodological Rigor: The introduction of vocabulary-controlled evaluation and the characterization of SPSA inversion rates address specific flaws in existing NLP benchmarks and optimization practices.

The authors conclude that while raw accuracy is modest compared to fine-tuned transformers, the results offer a "measurable, interpretable, causally identified structural signal" that connects Arabic grammatical tradition with quantum mechanics, laying the groundwork for future hardware experiments and morphological tensor architectures.