From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space

2026-03-13

AI Summary

This work addresses the modality gap between textual and time-series data, which hinders effective fusion of event-driven non-stationary information, as existing approaches struggle to reliably translate textual semantics into quantifiable predictive signals. To bridge this gap, the paper proposes the Temporal Evolving Semantic Space (TESS), which leverages large language models to extract structured temporal primitives—such as mean shifts, volatility, shape patterns, and lags—as cross-modal intermediate representations. A confidence-aware gating mechanism is introduced to suppress interference from redundant tokens. Evaluated on four real-world datasets, TESS significantly outperforms current unimodal and multimodal baselines, achieving up to a 29% reduction in prediction error while maintaining both high accuracy and interpretability.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.

Problem

• time-series forecasting
• text integration
• modality gap
• event-driven non-stationarity
• temporal semantics

Innovation

• Temporal Evolution Semantic Space
• modality gap
• time-series forecasting
• structured prompting
• numerical grounding