Enhancing Analysis of Earnings Calls: A Self-Supervised Approach to Extractive Summarization with ECT-SKIE

Abstract

Earnings conference calls are among the most significant sources of information regarding a firm's financial performance and strategic outlook. However, growing transcripts length make it difficult to manually analyze. This paper investigates the potential of ECT-SKIE, a self-supervised extractive summarization model, in addressing these challenges. This work systematically assesses the performance of ECT-SKIE in extracting key insights, the efficiency of ECT-SKIE compared with traditional methods, and the application of advanced techniques such as variational information bottleneck theory and structure-aware contrastive learning to improve the model's performance. Besides, the effectiveness of the container-based key sentence extractor in redundancy reduction is emphasized. A large-scale dataset of U.S. market earnings call transcripts is leveraged to verify the model with the ability of ECT-SKIE in significantly improving the accuracy, efficiency, and clarity of the extraction, making it a standard for automated financial analysis.