Motivation: Legacy rule-based expert systems still exist today, because they work well at solving a specific problem and deliver economic value for a business based on combined knowledge of human experts in the problem domain.

An example of a legacy rule-based system is one that offers discount payment incentives to debtors in order to entice them to repay. Characteristic of such legacy systems is that the knowledge used to arrive at the discount incentive is locked away, not only inside the case worker, but also the unstructured documents they produce.

Example unstructured documents include:

a) ticketing system documents: containing the expert’s rationale; instructions for handling accounts, and alternative strategies to offer debtors
b) legacy code: containing plain text rules
c) scanned letters: containing monetary incentive offered to a debtor and agreement conditions received from the debtor
d) email exchange with the debtor and case worker

Problem: Although a legacy system continues to bring value to the business they have several limitations:

1) Not Scalable: Legacy systems do not scale. The potential individuals that can be helped with the burden of debt is limited by the amount of cases a worker can process
2) Untune-able with Automated Optimized: It is not possible to experimentally optimize the amount of payment discount such that a favorable amount is offered to the debtor, yet minimize loss for the business
3) Lack Fairness: The bias of the case worker is baked into the discount. A human decides the terms and conditions of the discount and the system may be influenced by bias due to age or gender of debtor, season of the year, or human error.

Solution: We overcome the aforementioned limitations by modernizing a legacy, German-language rule-based system. We re-encode the rules of a rule-based system and use open source tools including: Tesseract and SpaCy to extract Machine Learning features from heterogenous documents (pdf, tickets, emails, etc). We present a case study which uses AI Fairness 360 to build tunable, Bias-Minimizing AI solutions that are capable of offering personalized, payment discounts to debtors.

In this talk you will learn techniques for building an Information Extraction pipeline from images and unstructured German text. You will also learn about the pitfall and successes from a case study to build a bias-mitigating AI solution for debt repayment.