Mining company sustainability reports to aid financial decision-making

Extracting information from financial documents like annual reports, sustainability reports or analyst reports plays an important role in investment decisions. Manual processing of these reports is time-consuming and tedious. In the past, pattern-based information extraction tools have been proposed to extract financial parameters from large documents. Rulebased approaches do not work for situations that need to extract information about actions or compliances along with targeted quantifiable information. In this paper, we present deep-learning based methodologies to retrieve information about sustainability practices from reports. Sustainability practices are becoming increasingly significant for investors to assess risk associated to a company. These reports have complex formats along with images and tables, due to which OCRs often fail to read the content correctly. We present methods to automatically detect text blocks from arbitrarily formatted PDF reports in a reliable way before using the OCR to read and index the content. This content is then searched for indicators that represent sustainability practices. The retrieved sentences are ranked based on their conceptual similarity to the indicators as well as quantitative content. Results show that the proposed methods retrieve sentences with high recall and precision and therefore substantially decrease human efforts to generate the right insights.