Your team spends hours each month tagging customer support tickets.
Whether it’s for tagging common questions, keeping track of unresolved bugs, or understanding how changes affect incoming queries, tagging is one of those very valuable features that we might take for granted. As Bryant Gillespie from shopVOX says on his blog, “Tagging is how you’ll know what types of conversations your users are having with your team. You will analyze those conversations you tag to improve your product and your customer service.”
Turning your qualitative customer conversations into quantitative data is incredibly helpful for bringing customer support closer to the product team.
Petr Pinkas at Feedly uses tags for an entirely different purpose. “I use them for follow-ups for bugs so I can let customers know when the bug was fixed” Petr shared with us. “Also when we implement a feature customers wanted for some time, I can let them know personally it is ready in our product, because I’ve tagged the tickets.”
With all this value coming from tags, wouldn't you want to make sure you’re doing it right?
Bryant shares how his team began using tagging in Intercom without any consistency or set process. “Maybe we saw it as something that we wouldn’t get immediate value out of. It just wasn’t a priority. Looking back - that was a dumb move.”
When you haven’t developed a structure of tagging, it’s hard to get any value out of historical data. Once you’ve closed the tickets, it’s a lot of work to go back and re-tag tickets, as Bryant found out.
“Eighteen months later, we closed over 10,000+ conversations. But we had diddly-squat to show for it” mourns Bryant. But that’s all changed for shopVox now (definitely check out his blog post on the process they went through!).
Inconsistent tagging affects your ticket analytics and your workflow. Not being clear with the structure of your tagging leaves agents confused (I once knew an agent who tagged every ticket as “general”). It means missing out on a lot of value.
To us, well-structured tagging is essential to training a machine learning algorithm. When we start working with new customer support teams, we begin by looking into their tagged data. The cleaner the data, the easier it is to train their machine learning model. The algorithm will be able to automate tagging with much more confidence when it’s fed data that can be easily digested by a computer.
So, for the sake of our algorithm, we’re passionate about learning how to tag correctly.
We give the same advice on tagging to every one of our clients, and we’d love to share it with you. Even if you aren’t looking to start using automation, clear and consistent tagging helps you understand what drives customer trends. What are customers most confused about? What type of tickets results in bad satisfaction surveys most often? Understanding these drivers is only possible through accurate tagging.
Here’s how your team can improve your tagging to make your data more effective (and ready for machine learning).
When tags don’t have a clear, unique definition it leads to confusion. Both humans and machines need really specific guidelines for what each tag should be used for. For example, if you create tags for both “issues” and “issues with data” there’s a clear overlap. Agents might use either tag for an issue relating to data, and there goes your consistency!
The best practice here is to avoid overlapping concepts and to define each of them. Try defining each of your tags with one or two sentences in a spreadsheet. If the sentence includes conditions like “except for X, which is filed with Y”, you might want to consider adjusting your tags.
Sometimes teams have created tags that only include a handful of tickets. They either thought the issue would arise often enough to require it, or haven’t revisited their tagging structure recently enough.
Having tags that don’t get used very often doesn’t work for training algorithms. The “big” tags will outweigh these tiny tags and create noise - leading to incorrectly labeled tickets. It doesn’t work for training humans either. Agents won’t remember these special tags exist if they don’t use them frequently. They won’t show up on trend reports because the percentage of tickets is too small compared to others, and it’s generally just a messy practice.
Instead, see if you can combine tags in a way that’s still meaningful. If you’re using this very specific tag for a very unique workflow, consider whether you’re actually making it simpler by requiring a tag. Review the long tail of tags frequently to prune the ones you don’t need anymore. While some teams insist that changing up their tagging scheme will lead to data problems, it’s actually better to make minor adjustments sooner rather than later. In this case, the tags are used so infrequently that it won’t affect your overall data analysis anyway.
We’ve seen CS teams with an huge number of tags defined (sometimes hundreds of them!). They are usually trying to account for every possibility. They want to make sure that if product asks “hey, did anyone write in about X”, they can jump in and pull out tickets. But this is the wrong approach. Due to human cognitive constraints, it’s almost impossible to tag consistently with that many tags. It’s also extremely time-consuming for agents to peruse through an enormous list of tags and choose one. It’s one of the reasons we strongly advocate for teams to automate customer support ticket tagging as soon as they have enough data.
When tagging, quality tops quantity.
Try to stick to a maximum of 30 tags. Your agents will find it easier to be consistent, and your data quality will thank you for it!
Most helpdesks will let you nest tags in a hierarchy so you can group subtags together. This helps your agents (and machines) sort through tags more easily.
As Raul explains in a previous post on creating text classifiers, machines are more easily trained with a well-structured tree. “Organize your tags according to their semantic relations. For example, Basketball and Baseball should be subtags of Sports because they are specific types of sports. Likewise, Clothing and Electronics should be a subset of tags of Retail. ”, Raul says. “A classification process that has a good structure can make a great difference and will be a huge help to make accurate predictions with your classifiers.”
Do a quick scan through your tags to see if there are any commonalities. If you can group any together, clean it up and make it much easier for your agents!
Okay, that heading is a bit jargony. But it basically means your tags shouldn’t compare apples to oranges. If you’re tagging based on product area, the reason for the contact and the fix applied, keep these as three different groups. In your help desk, I recommend separating each of these tags into custom fields. One field for product area (dashboard, billing, admin, etc), one field for the contact reason (bug, feedback, how-to), and one field for the fix (sent to engineering, updated account, sent to help center).
In machine learning, these groups of tags are called “models” and it’s an essential building block of training machines. But it’s important for humans too! Keeping these models separate helps avoid confusion and helps agents detect either/or conditions. In this example, they know that they need to apply at least three tags - one from each model.
If you’re hoping to automate in the future (and you should!), it’s important to examine your tags as a machine would. Machines are only able to look at the content stored in the ticket - they don’t have the context a human does.
Sometimes, CS teams resort to knowledge which is not explicitly stated in the ticket for tagging. For example, tagging tickets with client ids, email addresses, etc. This becomes a problem when training a machine learning algorithm. Machines do not have access to that knowledge and, therefore, cannot learn from it. It can actually lead to the algorithm making fake connections - and you might be unpleasantly surprised with its suggestions for tags going forward.
Instead, stick to using the information in the ticket for your tags, and integrate outside information in other ways.
Already following all of these classification rules? I think you’re ready to try creating a machine learning model for text classification.
As long as you have at least 20 tagged tickets per tag, your dataset is in good shape to get started. Plugging that data into a machine learning algorithm will produce consistent and accurate results. And the benefits of using machine learning are huge! Once you’ve created your machine learning model with MonkeyLearn (no coding required!) your team will never have to tag another ticket. Talk about saving your customer support team’s time.
Not there yet? Follow these guidelines to update your tagging strategies, and you’ll be ready to start automating in no time.
June 14th, 2018