Tapping into web-chat with Turn2Us

By Meg Leyland, Data Science Consultant and volunteer Data Ambassador with DataKind UK

The volunteers and staff at Starlight's DataDive weekend

"We were excited about using text analysis techniques to pull out recurring issues and topics, streamline interactions, and make finite resources stretch to even more people!"

Meg Leyland, project volunteer

In March, I was one of the three Data Ambassadors for a DataKind UK DataDive project with national financial help charity, Turn2Us. Here’s what we got up to…

Within the first half hour of our meeting with Turn2Us, we had a clear appreciation of two things: the vital need for the work that Turn2Us do supporting those in financial hardship, and the value of our opportunity to help. With a small army of data enthusiasts and a free weekend in March, we intended to give Turn2Us a great boost to their data journey.

In that first meeting, Michael Clarke, Head of Information Programmes, and Pip Johnson, Head of Insight and Impact at Turn2Us, introduced us Data Ambassadors (me, Ran, and Cal) to the work that Turn2Us does, including their website’s hugely popular online benefits calculator and grant search functionality, which have hundreds of thousands of visitors every month.

The impact on Turn2Us’s services during the pandemic

They also told us how thinly their services are being stretched, especially with the massive influx of users during the pandemic, which caused a 520% increase in demand for these services. At one point, 50,000 people per day used the online benefits calculator, and their Covid-19 grant had 5,000 applications in four days.

Of those, 92% had their employment affected, and 6 in 10 couldn’t afford food. Turn2Us also offers one-to-one advisors to complement their website’s services, but the phone line and web-chat are constantly busy, and the team also deals with 400 to 500 emails each week.

This is where we hoped to help. Turn2Us believed that there was an untapped wealth of information in their web-chat transcripts, through which people get support with their grants or benefits searches.

The goal was to paint a picture for Turn2Us of how people use their services and the common challenges raised by clients. By the end of our first chat, we were already very excited about the potential for using text analysis techniques to pull out recurring issues and topics, streamline interactions, and make finite resources stretch to even more people!

Over the next few weeks Ran, Cal and I worked to wrestle the data into shape so we could present it — with some carefully thought out guiding questions — to our intrepid weekend DataDivers. This process wasn’t without its challenges: as is so frequently the case with overstretched service delivery organisations, some of Turn2Us’s data was fragmented or gappy, and we sometimes had to decode mysteriously-designed databases set up by previous suppliers.

It was also really important to frame DataDive questions while keeping in mind how any potential insights could be used: it’s easy to get carried away with flashy data science, but we wanted to make sure we were delivering outcomes that Turn2Us could use to drive informed, transparent, ethical, and effective choices for their service delivery.

As the DataDive weekend kicked off with a Zoom call introducing us to the weekend’s volunteers, we were ready.

To those that haven’t heard of a DataDive before, I’d encourage you to imagine a marathon after-school coding club. There’s friendly faces, lots of fun chat, and (most importantly) bags of enthusiasm for data.

Over 48 hours, an astonishing volume of charts, dashboards, and insights get churned out, often coming from really creative approaches and unlikely collaborations, with the charity representatives on hand throughout to guide and give vital context.

As we talked through the findings to Michael and Pip throughout the weekend, there was a massive variety to showcase, such as:

• Dashboards to allow exploration of topics being talked about by demographic (Are certain benefits discussed more in the North than the South?)

• Algorithms to cluster users based on what they’re talking about

• Graphs to show the prevalence of certain topics over time (Are more people discussing ‘furlough’ now than before the pandemic?)

We were thrilled to see that a lot of this resonated with what they had already observed, meaning that transparent, data-driven decisions could be made based on what was previously just a hunch.

After the weekend, the Data Ambassadors reconvened to put a report together for Turn2Us, summarising the findings and offering recommendations for how to continue their data journey going forward. The three of us (all data scientists in our professional lives) agreed that our main recommendation would be a robust pipeline to turn messy web-chat transcripts into clean, usable data. This would elevate it from being html-riddled and completely unreadable, to analysable by off-the-shelf tools like PowerBI or Tableau, without straying too far into specialised data science.

Right from the start, Michael was keen to chase the incremental gains — small changes that, across the board, would make a big difference to their operations — and we thought this fitted the bill, as a manageable change to transform data they already have into meaningful insight.

I hope Turn2Us feel they’ve benefited from the findings of this project, and I’d really encourage anybody who thinks this project sounds interesting (either from a data or a charity perspective) to get Diving in the future!