Fin-EMRC: An Efficient Machine Reading Comprehension Framework for Financial Entity-Relation Extraction

Extracting entities and their relationships from financial documents is crucial for analyzing and predicting future market trends. However, the current state of the art in this field faces two major challenges: multiple sentences between related entities and poor few-shot performance caused by the vast amount of knowledge required in the financial domain. To address these challenges, we propose a framework for entity-relation extraction in financial documents that leverages multi-turn machine reading comprehension (MRC) and a Longformer model to handle long text dependencies. Furthermore, we propose a knowledge-enhanced method that incorporates structured knowledge from the financial domain. We also propose a new query template scheme that reduces the computational complexity of inferring complicated entity relations, thus improving the inference efficiency of the MRC framework. We evaluated our proposed framework on both general domain datasets and a real-world annotated Financial Entity-Relation (FinER) dataset. The results demonstrate that our Fin-EMRC framework outperforms currently available state-of-the-art methods in terms of entity and relation extraction accuracies. Moreover, the proposed efficient framework requires only 1.8 to 2.7 times less time for inference compared to the standard MRC framework.