An Exploratory Evaluation of LLM-Assisted Rewriting of Moderate-Complexity Financial Sentences for DisCoCat-Based Sentiment Analysis
Controlled LLM rewriting made harder financial sentences cheaper to process with DisCoCat, cutting circuit size by over 70%, but downstream accuracy improved only modestly.
The workflow compresses, simplifies, or splits financial sentences before DisCoCat processing. Its strongest variants cut average qubit and gate counts by **more than 70%**; GPT-4.1-mini with Prompt B reached **0.550 ± 0.035** mean accuracy.
Builders using agents as preprocessors should evaluate the transformed data against the downstream system, not just for linguistic fidelity. Prompt choice, filtering, and circuit cost all changed the result.
The workflow compresses, simplifies, or splits financial sentences before DisCoCat processing. Its strongest variants cut average qubit and gate counts by **more than 70%**; GPT-4.1-mini with Prompt B reached **0.550 ± 0.035** mean accuracy. Builders using agents as preprocessors should evaluate the transformed data against the downstream system, not just for linguistic fidelity. Prompt choice, filtering, and circuit cost all changed the result. The low-complexity baseline scored **0.521 ± 0.050**, so the observed accuracy gain was modest. Larger training splits were moderately associated with lower accuracy (**r=-0.446**), and the authors describe the findings as exploratory.
This confirms that LLM preprocessing should be treated as an optimization stage jointly evaluated with its downstream consumer. Large circuit-cost reductions did not translate into a comparably large accuracy gain, so prompt choice and filtering cannot be selected on linguistic plausibility or compression alone; the exploratory evidence also limits generalization beyond this financial DisCoCat workflow.