建構可靠的10-K財報可讀性衡量法－利用機器學習的文本清理減少可讀性中的雜訊

Abstract

隨著10-K文本分析的興起，如何準確地計算可讀性分數，以及哪種10-K文件的文本清理方法最有效，已變得至關重要。我們提出了基於機器學習的文本整理方法（LTT-Better），這個方法利用Bi-LSTM模型清理文本以利進行可讀性的計算。大多數可讀性公式假設文本內僅包含完整的句子，不包括標題或頁碼，很少原始10-K文件可以滿足這樣的條件。LTT-Better使用Bi-LSTM刪除10-K中不必要的字符，減少干擾，提高文本清理的品質。當使用LTT-Better代替傳統的文本清理時，大多數可讀性在統計上更接近人工清理的10-K報告。我們的研究進一步使用了1994年至2022年的10-K進行實證研究，調查可讀性引起的資訊不確定性是否能影響10-K提交日期後的股價波動。我們的實驗結果顯示，與傳統基於規則的文本清理相比，LTT-Better的可讀性在大多數情況下達到了更高的t分數。此外，當迴歸模型包含傳統文本清理的Fog指數和LTT-Better Fog指數時，兩者都具有顯著性，其中LTT-Better Fog指數的t分數更高。我們的研究結果顯示，當研究需要清理10-K報告以進行可讀性分析時，LTT-Better是一種有效的方法。未來的研究應在分析其語言特徵之前，將此清理方法應用於10-K文件。此外，我們向研究人員提供了關於使用不同文本清理方式後，應使用哪些可讀性公式的建議。

With the growth of 10-K text analysis, it becomes essential to determine how to reliably compute readability scores and what text preparation method for 10-K files is effective. We propose the Better Learning-Based Text Tidying (LTT-Better) approach that leverages Bi-LSTM models in preparing text for readability computation. Most readability measures assume correct sentence boundaries and text chunks without headings or dangling page numbers. These conditions are rarely satisfied in the original 10-K files. LTT-Better uses Bi-LSTM to remove unnecessary text chunks to reduce the noise and improve text preparation and text analysis using 10-K reports. When LTT-Better is used instead of the traditional rule-based preparation, the majority of the readabilities are shown to be statistically closer to the readabilities of human-prepared 10-Ks. Our research further conducts empirical models that investigate whether readability-induced information uncertainty can contribute to stock price volatility after the filing date using 10-Ks from 1994 to 2022. Our empirical results show that, compared to rule-based text preparation, readability from LTT-Better achieved a higher t-value in most cases. Moreover, when the regression models contain both the rule-based Fog index and LTT-Better Fog index, both are significant, with the LTT-Better Fog index achieving a higher t-value. Our findings suggest that LTT-Better is a promising approach to preparing 10-K reports for readability analysis. Future research should apply such an approach to 10-Ks before analyzing their linguistic attributes. Moreover, we give researchers helpful direction on what readability measurements should be used in future research.