Grammar Without Semantics: Sentiment Classification via Quantum Natural Language Processing

We investigate whether grammatical structure, independently of word meaning, encodes sufficient information to classify sentiment. To test this, we use Quantum Natural Language Processing (QNLP) as a controlled probe: by fixing all gate parameters in a DisCoCat- derived quantum circuit to a single constant angle, we obtain a feature extractor that captures syntactic structure while discarding all lexical semantics. Applied to a dataset of over 4,500 prompts labeled as negative or non-negative, the grammar-only extractor achieves an AUC of ≈ 0.80. This supports the hypothesis that systematic co-occurrences between grammatical constructions and sentiment labels are present and learnable at corpus level. The primary limitation of all quantum approaches is inference time, which exceeds the classical reference by one to three orders of magnitude–the parsing of sentences into diagrams rather than the quantum component itself being the main bottleneck.