Interpretability of deep learning models in analysis of Spanish financial text

Abstract

Artificial intelligence methods based on deep learning (DL) have recently made significant progress in many different areas including free text classification and sentiment analysis. We believe that corporate governance is one of these areas, where DL can generate very valuable and differential knowledge, for example, by analyzing the biographies of independent directors, which allows for qualitative modeling of their profile in an automatic way. For this technology to be accepted it is important to be able to explain how it generates its results. In this work we have developed a six-dimensional labeled dataset of independent director biographies, implemented three recurrent DL models based on LSTM and transformers along with four ensembles, one of which is an innovative proposal based on a multi-layer perceptron (MLP), trained them using Spanish language and economics and finance terminology and performed a comprehensive test study that demonstrates the accuracy of the results. We have also performed a complete study of explainability using the SHAP methodology by comparatively analyzing the developed models. We have achieved a mean error (MAE) of 8% in the modeling of the open text biographies, which has allowed us to perform a case study of time analysis that has detected significant variations in the composition of the Standard Expertise Profile (SEP) of the boards of directors, related to the crisis of the period 2008–2013. This work shows that DL technology can be accurately applied to free text analysis in the finance and economic domain, by automatically analyzing large volumes of data to generate knowledge that would have been unattainable by other means.