Scholarlink

<p>Extracting structured causal knowledge from complex financial narratives (Event Causality Extraction, ECE) is pivotal for intelligent market analysis. However, existing methods often treat event labels as abstract indices, ignoring their intrinsic semantic definitions. Consequently, these models struggle to distinguish semantically ambiguous causal types (e.g., Cost-driven vs. Price-driven). Furthermore, while Large Language Models (LLMs) have shown promise, they suffer from high inference latency and often lack the boundary precision required for fine-grained argument extraction. To address these challenges, we propose DisPrompt, a novel joint extraction framework characterized by discriminative semantic prompting. By integrating Supervised Contrastive Learning (SCL) with a Semantic Label Attention Network (LSAN), we transform extraction into a precise, semantics-driven matching process while enforcing rigorous discrimination between ambiguous causal logics and design a Deep Co-Attention Fusion (DCAF) module that explicitly captures the structural interplay between cause-and-effect arguments. Experiments on the Chinese ECE-CCKS benchmark show that DisPrompt achieves a 62.02% F1 score, outperforming the previous state-of-the-art by 7.00%. Extended experiments on the English FinCausal dataset demonstrate strong cross-lingual generalization, surpassing GPT-4-based retrieval-augmented baselines by 10.2%.</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

DisPrompt: joint extraction of financial causal events via discriminative semantic prompting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月26日

查看全文
  • Nietong Wang,
  • Yinhua Tian,
  • Fangyu Cheng,
  • Chen Shi,
  • Cong Liu

摘要

Extracting structured causal knowledge from complex financial narratives (Event Causality Extraction, ECE) is pivotal for intelligent market analysis. However, existing methods often treat event labels as abstract indices, ignoring their intrinsic semantic definitions. Consequently, these models struggle to distinguish semantically ambiguous causal types (e.g., Cost-driven vs. Price-driven). Furthermore, while Large Language Models (LLMs) have shown promise, they suffer from high inference latency and often lack the boundary precision required for fine-grained argument extraction. To address these challenges, we propose DisPrompt, a novel joint extraction framework characterized by discriminative semantic prompting. By integrating Supervised Contrastive Learning (SCL) with a Semantic Label Attention Network (LSAN), we transform extraction into a precise, semantics-driven matching process while enforcing rigorous discrimination between ambiguous causal logics and design a Deep Co-Attention Fusion (DCAF) module that explicitly captures the structural interplay between cause-and-effect arguments. Experiments on the Chinese ECE-CCKS benchmark show that DisPrompt achieves a 62.02% F1 score, outperforming the previous state-of-the-art by 7.00%. Extended experiments on the English FinCausal dataset demonstrate strong cross-lingual generalization, surpassing GPT-4-based retrieval-augmented baselines by 10.2%.