ECTSum

尽管在自动摘要方面取得了巨大进步，但先进的方法主要经过培训，可以在总结简短的newswire文章或具有强烈版面偏见的文档 (例如科学文章或政府报告) 方面表现出色。总结财务文件 (包括事实和数据) 的有效技术在很大程度上尚未得到探索，主要是由于没有合适的数据集。在这里，我们介绍了ECTSum，一个新的数据集，其中包含由上市公司托管的收益电话 (ECTs) 的笔录，作为文档，以及简短的专家撰写的telegramstyle子弹头摘要，这些摘要源自相应的路透社文章。ECTs是长的非结构化文档，没有任何规定的长度限制或格式。我们通过各种指标使用最先进的摘要器对ECTSum进行基准测试，以评估生成的摘要的内容质量和事实一致性。最后，我们提出了一种简单有效的方法ECT-BPS，以生成一组项目符号点，以精确捕获通话中讨论的重要事实。

Despite significant advances in automatic summarization, state-of-the-art methods are primarily trained to excel at summarizing short newswire articles or documents with well-defined structural conventions (e.g., scientific papers or government reports). Effective techniques for summarizing financial documents—including those containing factual information and numerical data—have largely remained unexplored, primarily due to the absence of suitable datasets. Here, we introduce ECTSum, a novel dataset consisting of earnings conference call (ECT) transcripts from publicly traded companies as source documents, paired with concise Telegram-style bullet-point summaries expertly crafted based on corresponding Reuters articles. ECTs are lengthy, unstructured documents with no preset length limits or standardized formatting. We benchmark ECTSum using state-of-the-art summarizers across multiple metrics to evaluate the content quality and factual consistency of the generated summaries. Finally, we propose ECT-BPS, a simple yet effective method to generate bullet-point summaries that accurately capture the key important facts discussed during the earnings calls.

搜集汇总

背景与挑战

背景概述

ECTSum是一个针对财务文档摘要的数据集，包含上市公司收益电话会议（ECTs）的转录文本和专家撰写的要点摘要，用于评估摘要模型的内容质量与事实一致性。该数据集由汉诺威莱布尼兹大学等机构于2022年发布，旨在填补财务文档摘要领域的数据空白。