VerbalTS: Generating Time Series from Texts

Shuqi Gu, Chuyue Li, Baoyu Jing, Kan Ren

TL;DR: VerbalTS is a framework that generates time series from unstructured text. It uses multi-focal alignment for improved quality and semantic match, outperforming existing approaches on various datasets.

Abstract: Time series synthesis has become a foundational task in modern society, underpinning decision-making across various scenes. Recent approaches primarily generate time series from structured conditions, such as attribute-based metadata. However, these methods struggle to capture the full complexity of time series, as the predefined structures often fail to reflect intricate temporal dynamics or other nuanced characteristics. Moreover, constructing structured metadata requires expert knowledge, making large-scale data labeling costly and impractical. In this paper, we introduce VerbalTS, a novel framework for generating time series from unstructured textual descriptions, offering a more expressive and flexible solution to time series synthesis. To bridge the gap between unstructured text and time series data, VerbalTS employs a multi-focal alignment and generation framework, effectively modeling their complex relationships. Experiments on two synthetic and four real-world datasets demonstrate that VerbalTS outperforms existing methods in both generation quality and semantic alignment with textual conditions.

Lay Summary:

Imagine you’re planning a vacation and want to understand how the weather might change over time. Or maybe you’re a traffic manager trying to figure out how traffic flows through a city during the day. To do this, people often rely on time series—patterns that show how things change step by step. But creating these patterns is tricky. Most existing tools require experts to provide structured labels, which is slow, expensive, and often misses important details.
So we asked: what if we could just verbally describe what we want in natural language—like “a sudden drop, then a slow rise”—and have a computer bring that pattern to life? That’s exactly what our new system, VerbalTS, does. It takes simple, natural descriptions and turns them into realistic patterns that behave like the real thing. We tested it across many different scenarios, and it consistently outperformed other tools. This could make it easier for people in many fields to explore “what if” questions, without needing to be data experts.

Primary Area: Applications->Time Series

Keywords: time series, conditional generation, diffusion model, multi-modality

Submission Number: 3470