Risk-Aware Automatic Text Summarization with Pre-Identification of Risk Categories and Emphasis

Automatically generating structured analysis from a financial document is both a critical and challenging task. Unlike the simple document summarization typically performed by large models, structured analysis not only extracts key information but also involves deeper categorization, interpretation, and mapping of relationships. The challenge arises from the model's need to uncover relationships between data points, assess potential financial impacts, and forecast trends, all of which require a sophisticated understanding of the content and context. In this paper, we introduce ATS-PRCE, a novel approach that utilizes LLMs to automatically generate analyses with higher precision and pertintness, thereby improving the model's understanding and analysis quality of input text. ATS-PRCE first prompts LLMs to identify potential risk types in financial documents and then guides LLMs to build structured analysis according to the identified risk types. We conducted extensive experiments on the benchmark dataset, gretel-financial-risk-analysis-v1, utilizing five widely used and powerful LLMs. Our findings demonstrate that ATS-PRCE achieves notable and consistent improvements acorss all LLMs and benchmarks. For example, ATS-PRCE improves the Similarity metric by 7% over the GPT-3.5 model.