Extraction of Numerical Facts from German Texts to Enrich Internal Audit Data

Large-scale automated data processing is usually only possible for internal auditors in the case of structured data. Unstructured data, such as facts contained in texts, on the other hand, are often processed manually and using sampling. This, in turn, can increase the risk of disregarding relevant information during an audit. To address this risk, we present an approach that can be used to extract numerical facts along with their associated entities and relations from German texts and convert them into a format that can be processed by audit tools. The algorithm developed for this purpose follows a rule-based logic and was evaluated using 4637 sentences from 50 German annual reports. The results show that in more than 75% of all cases, the entity and relation of a numeric value within the sentence could be determined correctly.