Abstract
Time-series Generation (TSG) is a prominent research area with broadapplications in simulations, data augmentation, and counterfactual analysis.While existing methods have shown promise in unconditional single-domain TSG,real-world applications demand for cross-domain approaches capable ofcontrolled generation tailored to domain-specific constraints andinstance-level requirements. In this paper, we argue that text can providesemantic insights, domain information and instance-specific temporal patterns,to guide and improve TSG. We introduce ``Text-Controlled TSG'', a task focusedon generating realistic time series by incorporating textual descriptions. Toaddress data scarcity in this setting, we propose a novel LLM-based Multi-Agentframework that synthesizes diverse, realistic text-to-TS datasets. Furthermore,we introduce BRIDGE, a hybrid text-controlled TSG framework that integratessemantic prototypes with text description for supporting domain-level guidance.This approach achieves state-of-the-art generation fidelity on 11 of 12datasets, and improves controllability by up to 12% on MSE and 6% MAE comparedto no text input generation, highlighting its potential for generating tailoredtime-series data.