Author: Nick Masca
Here at Resolver Group, we want to help businesses and consumers when things go wrong by amplifying the consumer voice and making complex complaints data actionable.
Given the impending public health crisis due to the spread of the Covid-19 coronavirus, we’ve decided to open up a capability that we believe can help companies better support those customers most in need.
From today, we’ll be offering free access to our Egeria Analytics: Vulnerability Assistant portal, enabling companies to identify situationally vulnerable consumers who have submitted a complaint via our resolver.co.uk channel, and thereby provide the help and support these customers need as a matter of priority.
We believe our tool can help businesses provide much needed support to consumers in vulnerable situations, who may be struggling to get in touch via other channels.
The Vulnerability Assistant detects ‘triggers’ of up 16 types of situational vulnerability from complainants’ case text, leveraging our advanced natural language processing (NLP) capabilities.
As demonstrated in the report below, the Vulnerability Assistant flags cases where a vulnerability has been detected, providing details of the case (responses to the questions: ‘What Happened?’, ‘How Did This Affect You?’, and ‘What do you Want?’) and the contact details of the customer. This enables complaints handlers and/or customer service managers to identify their most at risk customers and quickly triage the set of complainants.
The Vulnerability Assistant uses Natural Language Processing techniques to transform complaint text into a numeric form that can be processed by machine learning software. We applied ‘supervised learning’ algorithms to individual ‘vectorised’ sentences from complaints, harnessing a rich dataset we built up over time by our in-house team of expert annotators.
Our annotators were trained to recognise signs (or ‘triggers’) of situational vulnerability in complaint text, providing us with over 70,000 labelled sentences to learn from. As vulnerability can be somewhat difficult to define and has an element of subjectivity, we used an iterative cycle of annotation and review, requiring agreement between multiple annotators to ensure high quality labels.
Like all machine learning algorithms, we cannot claim that ours perform perfectly. However, the ‘precision’ of our algorithms - that is, the chance that a positive vulnerability flag actually shows signs of a vulnerability, averages more than 75%, which we believe can significantly enhance complaints handlers’ abilities to identify and action high priority cases.
We motivated our choice and definitions of situational vulnerability types on guidance from regulators such as the Financial Conduct Authority (FCA), initially focussing on the following vulnerabilities:
While our current tool identifies situationally vulnerable consumers who have raised complaints via the resolver.co.uk website, we have the capability to also help identify vulnerabilities from other text too - such as social media posts, webchats or call-centre transcripts.
If you’d like to trial our existing product or discuss how else we can help with the support of vulnerable customer speak directly with our lead consultants who can arrange the set-up of this.