The Japanese Society for Artificial Intelligence

Comprehending the causal relationships among economic events is crucial for risk management because it aids in forecasting potential external shocks and formulating informed predictions regarding the results of prospective actions. The recent advancement of large language models (LLMs) offers a viable method for extracting domain-specific knowledge from textual content to develop causal graphs. Nonetheless, accurately identifying causal relationships that align with expert evaluations remains challenging in computational text analysis, particularly for financial and economic documents that demand specialized expertise. In response to this issue, we introduce a method utilizing causal-sensitive sentence embeddings, which excel in discerning causal relationships through fine-tuning text embedding models employing contrastive learning. This method employs a weakly supervised learning paradigm, generating the necessary training dataset for contrastive learning from extensive textual corpora via causal cues and LLMs. The evaluation experiments on four datasets against baseline methods highlight the effectiveness of our method.