TICK: Text-Informed Contrastive Kernels for Probabilistic Time-Series Retrieval

The coexistence of sensor data and textual knowledge across industrial applications highlights the need to align natural language with multivariate time series. This challenge is typically treated as a deterministic task, ignoring the fact that a vague query may match hundreds of recordings while a precise one matches only a few. We argue that text-to-time-series retrieval is inherently a probabilistic cross-modal problem and propose TICK, a framework in which each text query carries a learned uncertainty $\sigma_t$ reflecting how many time series are semantically compatible with it. Without supervising $\sigma_t$ directly, we find that it tracks query specificity with Spearman correlations $\rho \in [-0.88, -0.75]$ on three independent industrial benchmarks. We further show that coupling the time-series representation to the text query is a necessary condition for this calibration to emerge: without it, $\sigma_t$ collapses to a constant, eliminating the model's ability to adapt its retrieval scope to query specificity. Because $\sigma_t$ sets the retrieval scope of a query, TICK spends a query-dependent candidate budget instead of a fixed worst-case one, on top of a frozen 33M-parameter sentence encoder and a small patch Transformer, with no language model in the inference loop.