TLDR: A new framework called NUM2EVENT enables large language models (LLMs) to infer interpretable, structured events directly from numerical time-series data, even when textual context is unavailable. It uses an agent-guided extractor for event definitions, a synthetic data generator to overcome data scarcity, and a two-stage finetuning process to align numerical patterns with semantic event hypotheses, significantly outperforming existing LLM baselines in event prediction.
Large language models (LLMs) have shown remarkable abilities in understanding and reasoning with various types of data, especially text. However, their capacity to interpret purely numerical time-series signals, particularly for uncovering the underlying events that drive changes, has been quite limited. Traditional approaches often focus on forecasting numbers or describing trends, without explaining the ‘why’ behind these numerical shifts or mapping them to human-understandable events.
This limitation is particularly critical in fields like financial risk management and industrial operations. In these high-stakes environments, numerical anomalies often appear before any textual reports or news, and timely decisions are paramount. Existing methods struggle to provide explicit predictions of these underlying events when text is absent, making it difficult to respond quickly to situations like sudden market shifts or system faults.
Introducing NUM2EVENT: Bridging Numbers and Events
To address this significant gap, researchers have introduced a novel task: number to event reasoning and decoding. This aims to infer interpretable, structured events directly from numerical inputs, even when current textual context is unavailable. The new framework, called NUM2EVENT, is designed to explicitly reason over numerical changes, generate intermediate explanations, and output structured event hypotheses.
The NUM2EVENT framework integrates several key components to achieve this:
- Agent-Guided Event Extractor (AGE): This module acts like an intelligent agent, extracting structured events (defined by Actor, Action, Object, Direction – AAOD slots) from historical text. It builds an extensible vocabulary of events and ensures that the extracted information is standardized and consistent, providing high-quality supervision for the model.
- Event-Driven Time-Series Generator (EveDTS): Real-world data linking numerical time-series with structured events is scarce. EveDTS tackles this by synthesizing realistic data. It uses Hawkes processes to model how events arrive and influence each other, and local projection impulse responses to simulate how these events impact numerical series. This synthetic data helps the model learn diverse dependency patterns and impact effects.
- Two-Stage Finetuning Pipeline: The training process is divided into two stages. First, a dedicated time-series encoder is trained to learn robust representations from raw numerical segments. In the second stage, with the encoder parameters frozen, a large language model is finetuned. This allows the LLM to effectively align these numerical representations with event semantics, enabling it to decode structured events directly from numerical inputs.
By combining these components, NUM2EVENT enables the model to not only understand numerical fluctuations but also to translate them into meaningful, interpretable event hypotheses. This means the AI can explain and predict events directly from the dynamics of numerical data.
Also Read:
- Decoding Numbers: How Language Models Internally Process Numerical Information
- Unlocking Black-Box Optimization: How GPTOpt Leverages LLMs for Efficiency
Performance and Impact
Experiments conducted on real-world datasets from domains like Energy (U.S. gasoline prices) and Public Health (Influenza-Like Illness cases) demonstrate the effectiveness of NUM2EVENT. The model substantially outperforms strong LLM baselines in event-level precision and recall. For instance, in the Energy dataset, it achieved more than double the precision and recall of the strongest baseline, highlighting its superior ability to bridge quantitative reasoning and semantic event understanding.
Ablation studies further confirmed the importance of each component within the NUM2EVENT framework, showing that removing any part led to a consistent degradation in performance. This research opens a new direction for AI, allowing large language models to move beyond mere forecasting or trend description to provide causal and semantic reasoning about real-world events directly from numerical signals. For more technical details, you can refer to the full research paper here.
