@inproceedings{wang-etal-2025-omnieval,
    title = "{O}mni{E}val: An Omnidirectional and Automatic {RAG} Evaluation Benchmark in Financial Domain",
    author = "Wang, Shuting  and
      Tan, Jiejun  and
      Dou, Zhicheng  and
      Wen, Ji-Rong",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-emnlp/2025.emnlp-main.292/",
    pages = "5737--5762",
    ISBN = "979-8-89176-332-6",
    abstract = "Retrieval-augmented generation (RAG) has emerged as a key application of large language models (LLMs), especially in vertical domains where LLMs may lack domain-specific knowledge. This paper introduces OmniEval, an omnidirectional and automatic RAG benchmark for the financial domain, featured by its multi-dimensional evaluation framework: First, we categorize RAG scenarios by five task classes and 16 financial topics, leading to a matrix-based structured assessment for RAG evaluation; Next, we leverage a multi-dimensional evaluation data generation method that integrates GPT-4-based automatic generation and human annotation approaches, achieving an 87.47{\%} acceptance ratio in human evaluations of generated instances; Further, we utilize a multi-stage evaluation pipeline to assess both retrieval and generation performance, resulting in an all-sided evaluation of the RAG pipeline. Finally, rule-based and LLM-based metrics are combined to build a multi-dimensional evaluation system, enhancing the reliability of assessments through fine-tuned LLM-based evaluators. Our omnidirectional evaluation experiments highlight the performance variations of RAG systems across diverse topics and tasks and reveal significant opportunities for RAG models to improve their capabilities in vertical domains. We open source the anonymous code of our benchmark at https://github.com/RUC-NLPIR/OmniEval."
}