Evaluating Tabular and Textual Entity Linking in Financial Documents.

Entity linking poses a longstanding challenge within natural language processing, a challenge that many studies have sought to address from diverse perspectives. However, these solutions have focused on open-domain knowledge bases. In contrast, our study shifts focus towards the financial domain, characterized by complexities, including intricate concept definitions and tables with complicated hierarchical structures. We introduce a novel dataset comprised of a hybrid mixture of tabular and textual inputs for entity linking, providing a total of 10,000 samples. The dataset requires predicting a concept that describes a specific table cell from the given table and surrounding texts. Our experiment includes three approaches. The first one is an approximate string-matching algorithm. The second is a fine-tuned language-model retriever. The last one involves prompting on a large language model. The results highlight the challenges inherent in this task, suggesting the need for further solutions.