Many situations require large organizations to actively monitor media for risk exposure to stakeholder groups. For example, anti-money laundering (AML) regulation requires banks to screen account holders when opening an account and on an ongoing basis. Journalists monitor media (including social media) to obtain leads for news stories. Operational risk managers, hedge funds and investment managers want to know the risk exposure associated with their current investment portfolio and potential new investment targets. Therefore, many interested parties have started to support their manual monitoring for risk exposure to companies or people with automation support, e.g. using text alerts based on keyword lists with relevant words or phrases ('slavery', 'corruption', 'forced labor') to obtain business advantages or to comply with regulations more effectively.
However, keyword-based methods either do not fully account for the variability of language (if they are incomplete) or they create a lot of spurious matches due to a lack of understanding (if they are exhaustive), the so-called false positive problem. For example, “fine” could denote a regular enforcement action in one context, yet in another context denotes happy emotional state; the latter instances lead to erroneous alerts (false positives) by string matching technology.
In this paper, we give a definition of the problem of computer-supported risk identification for people and company entities. We then demonstrate a proposed solution for the false positive problem, which is based on a combination of taxonomy learning, natural language processing and machine learning. We also demonstrate how we our method can learn new risk language on an ongoing basis (new future risk types lead to new names for these risks) in order to avoid our system to go ‘stale’. Seeral use cases for applications of the technology are outlined, and we conclude discussing some remaining technical challenges the privacy implications of this new technology.