Commonwealth Bank of Australia (CBA) is preparing to initiate a trial program in which instances of abusive language within transaction descriptions will be referred to the New South Wales (NSW) Police, contingent upon the consent of the victim.
CBA’s Pilot Program for Reporting Transaction Abuse to NSW Police The bank has announced that if its machine-learning model detects offensive language in transactions received by NSW customers, it will seek permission from the recipient to report the incident to the police.
CBA’s detection model has been scanning transaction descriptions received by its customers since 2021. This capability was developed by CBA’s AI Labs team following an incident in 2020 where over 8000 customers faced harassment through small-value deposits to their bank accounts.
According to CBA, their model has identified nearly 400,000 offensive language-containing transactions from CommBank app and NetBank users, attributed to around 1500 individuals on an annual basis.
Transactions flagged by the model undergo manual review by CBA’s anti-financial abuse team, Next Chapter. Depending on the situation, Next Chapter can block messages, suspend perpetrators from digital banking for three months, or even impose permanent suspension for repeated violations of the bank’s acceptable use policy.
CBA has disclosed that Next Chapter will initiate a pilot program starting in mid-September, wherein recipients of abusive messages will be given the option to have the bank file a police report on their behalf.
Angela Macmillan, CBA’s Group Customer Advocate, stated that this collaboration with NSW Police demonstrates the bank’s commitment to supporting victims and preventing abuse.
Distinguishing Feature of the Initiative While other major banks have also employed advanced analytics and algorithms to identify and prevent harassment through low-value transaction descriptions, CBA’s approach enables a streamlined and seamless request for police intervention, setting it apart as an innovative initiative.
Macmillan emphasized the groundbreaking nature of this collaboration between the banking industry and law enforcement, expressing hopes that it will pave the way for more effective cooperation in combating domestic and financial abuse.
Technical Aspects of CBA’s Model CBA Senior Data Scientist Dr. Anna Leontjeva, a key contributor to the project, explained that perpetrators resort to sending low-value transactions because they have been blocked on other platforms and are seeking alternative means to communicate.
To detect abusive transaction descriptions, CBA’s model considers various factors at both the transaction and relationship levels. Specifics such as dollar amounts, frequency, and textual analysis of the description field are assessed, along with variables like text length, case usage, word count, presence of special characters, and more.
Three trained language models analyze emotion, toxicity, and sentiment within the descriptions. These findings are then aggregated to evaluate the relationship between a potential abuser and victim. If an abuser has multiple victims, this is identified as distinct high-risk relationships.
Ultimately, a random forest model classifies the relationships as “highly abusive” or “non-abusive.”
CBA’s initiative showcases a proactive approach to tackling transaction-related abuse and sets a precedent for collaborative efforts between financial institutions and law enforcement agencies.
