Scholarlink

<p>Event causality extraction in the financial domain is crucial for understanding complex cause-effect chains behind market movements and economic events. However, traditional event relation mining often focuses on simple, single-cause links and struggles with the multi-cause, multi-effect scenarios frequently seen in financial news. This paper addresses financial event multi-causal correlation mining, a task that entails identifying all cause-effect event pairs in text, including direct and indirect causal influences, with structured event information. We formally define the task and highlight key challenges: the need to extract structured events (e.g. actor, action, object, time, location) from complex financial narratives, and to discern intertwined causal relations that may span multiple sentences or intermediate events. Existing methods often remain limited to surface-level associations and lack deep reasoning, leading to incomplete causal understanding. To overcome these issues, we propose a novel Large Language Model (LLM)-augmented framework, leveraging a fine-tuned LLaMA backbone as a powerful contextual encoder while integrating structured reasoning modules. The LLM is used to enrich event representations and capture domain-specific context, and we design an event-centric graph reasoning component that identifies direct causal links and infers indirect causality through multi-hop propagation. Our methodology enables comprehensive extraction of causal chains in financial texts. Experiments on three public benchmarks that our model significantly outperforms ten baseline approaches, including both classical models and other LLM-based methods. Notably, our approach achieves up to 5–10% F1 improvements in detecting complex multi-causal relationships. Further ablation studies confirm the effectiveness of each component, and case analyses illustrate how the LLM captures subtle causal cues (e.g. economic indicators triggering market reactions) that simpler models miss. These results underscore the value of combining LLMs with structured causal reasoning for high-precision event causality mining. The proposed research provides a more accurate and efficient tool for financial analysts to automatically uncover hidden causal links, supporting better decision-making and risk assessment in financial domains.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-enhanced multi-causal event causality mining in financial texts

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月11日

查看全文
  • Chunyu Yan

摘要

Event causality extraction in the financial domain is crucial for understanding complex cause-effect chains behind market movements and economic events. However, traditional event relation mining often focuses on simple, single-cause links and struggles with the multi-cause, multi-effect scenarios frequently seen in financial news. This paper addresses financial event multi-causal correlation mining, a task that entails identifying all cause-effect event pairs in text, including direct and indirect causal influences, with structured event information. We formally define the task and highlight key challenges: the need to extract structured events (e.g. actor, action, object, time, location) from complex financial narratives, and to discern intertwined causal relations that may span multiple sentences or intermediate events. Existing methods often remain limited to surface-level associations and lack deep reasoning, leading to incomplete causal understanding. To overcome these issues, we propose a novel Large Language Model (LLM)-augmented framework, leveraging a fine-tuned LLaMA backbone as a powerful contextual encoder while integrating structured reasoning modules. The LLM is used to enrich event representations and capture domain-specific context, and we design an event-centric graph reasoning component that identifies direct causal links and infers indirect causality through multi-hop propagation. Our methodology enables comprehensive extraction of causal chains in financial texts. Experiments on three public benchmarks that our model significantly outperforms ten baseline approaches, including both classical models and other LLM-based methods. Notably, our approach achieves up to 5–10% F1 improvements in detecting complex multi-causal relationships. Further ablation studies confirm the effectiveness of each component, and case analyses illustrate how the LLM captures subtle causal cues (e.g. economic indicators triggering market reactions) that simpler models miss. These results underscore the value of combining LLMs with structured causal reasoning for high-precision event causality mining. The proposed research provides a more accurate and efficient tool for financial analysts to automatically uncover hidden causal links, supporting better decision-making and risk assessment in financial domains.