This article is written by our in-house NLP scientist, Suhan Prabhu. In this article, Suhan explains how NLP support ticket tagging works and why that means our AI-powered Zendesk integration for support ticket insight is so powerful.
There's a vast quantity of customer insight hidden away in your support ticket logs.
If you unlock that data, you unlock the potential for support-driven growth.
Unlike other sources of customer feedback, support tickets are rich, unbiased, and come in high volumes.
In other words, they're a goldmine for businesses wanting to get closer to their customer, improve customer satisfaction, and tackle churn.
Manual support ticket analysis is not a realistic option. It's complex and subjective, making it time consuming and inaccurate to understand reasons for contact at a meaningful scale.
This is where AI comes into its own. The latest advances in natural language processing (NLP) make it actually a trustable source of customer care automation.
With NLP, you can now accurately understand the drivers of hundreds of thousands of tickets in real-time.
These kind of real-time analytics are essential to the future of customer service—automated ticket routing, prioritsation of tickets that drive churn, identifying growth opportunities in conversations, and anomaly alerting.
NLP makes uncovering detailed insight from high volumes an easy task. That data insight and how that insight is used is why in the next 3-5 years we believe customer service will be considered the most strategically useful business function.
In this article, we'll discuss what NLP is and why NLP technology is the foundational layer behind this future.
Natural language processing has recently emerged as a popular subfield of computer science and artificial intelligence (AI) that applies software to automatically manipulate natural language such as speech and text.
NLP systems can be one of two types:
Let's take an example of how each would handle three customer support tickets:
Ticket one: Hi, my delivery was never recieved. Please help me find it.
Ticket two: Hi, I'm here to complain about my delivery. It hasn't arrived yet.
Ticket three: Hi, my box was due to arrive between 5-7pm yesterday. I woke up this morning and it still has not been delivered
As humans, we know these three tickets have equal meaning. The box was not delivered.
A rule-based system built for a company wanting to track how many failed deliveries they had would be built with a rule like: Tag the ticket if it includes the words 'box', 'not', and 'delivered' anywhere within it. It would only tag ticket three correctly, a 66% failure rate.
A machine learning-based system, however, would correctly tag all three. It would understand that sometimes people use 'arrived' instead of delivery, and 'received' instead of delivered. It would also understand that, in ticket two, when the customer says 'it', they are referring to the box.
ML based systems are generally more robust when it comes to analysing misspelled words or ones that haven't been seen before. A rule based system would need a lot of rules to capture all three of these tickets, and output would need explicit error handling, which is time-consuming and difficult.
Customer support tickets are growing to be one of the main sources of customer data in companies. Support tickets can contain tickets raised in social media chats, email conversations and web-chat based conversations, and many companies receive them in the thousands everyday. They're also a rich, qualitative form of customer feedback, and each ticket contains information related to friction between the customer and your product, operations, and partners.
Support ticket conversations are, however, very raw, unstructured and unpredictable in length. Many of them are agent-customer conversations, some are complaints and each customer has a different ways of expressing themselves. Going through such large volumes of tickets requires considerable manual effort and, more often than not, important topics are missed. Even simplistic NLP (rule-based) struggles with support ticket analytics, missing insights due to the frequent presence of misspelled words and grammatically incorrect sentences.
Zendesk, the popular help desk platform, helps companies organize ticket requests, bringing them together on a single comprehensive dashboard. It makes use of a naive rule-based system to tag and categorize tickets. However, these tags are high level and coarse such as "website_issue", "presale_enquiry" and so on. They do not mention granular reason about the reason for customer contact. Any analysis still takes significant data handling to get accurate insights, and even then, they're likely to be high level or somewhat subjectiver.
Machine learning can bring something special to Zendesk: hierarchy. Hierarchy means that, when tickets are AI tagged, the relationship between ticket tags is noted. For example, lets consider a meal kit delivery company receiving a ticket from a customer about a missing ingredient. Zendesk's tagging system just tags this ticket as "objection raised" or perhaps, if you're lucky, "missing ingredient". But, it does not contain the finer details such as which ingredient is missing.
Looking at an individual conversation, that's not a major issue because we can read that it's about rice. However, in the thousands of tickets, going through all of those in the "missing ingredient" pile every day becomes complex and time-consuming. A properly built machine learning system would tag the ticket with Box Contents--> Missing Ingredient --> Rice, which would allow you to choose the level you'd like to analyse at and dig deeper to the root cause when you need to.
This also goes for multiple topics. Most support tickets contain 2-3 complaints and a rule-based system may only apply one topic to it.
Without this level of granularity at scale, you have no way of understand which customer issues are generating the most support tickets. And, no way of knowing their business impact.
We at SentiSum, have built a robust Zendesk chat analytics integration using the latest developments in machine learning-based NLP. Our system will analyze and categorize high volumes of support tickets in a matter of minutes. Our ML-based system accurately uncovers multi-level topic tagging, including general (Box Contents), granular (Missing Ingredients), and specific aspects (Rice). On applicable tickets, like surveys and customer review data, we also uncover the customer sentiment toward the topic raised.
Let us take you through how this works:
Once you have NLP support ticket insights in an actionable format, you unlock a whole host of business benefits.
The first is the most obvious, you are able to share the insight into customer problems with the team who can change the problem's cause. Whether it's a product issue or an operations issue, passing them actionable insights gives them a chance to make amends.
Integration our NLP analytics into Zendesk also unlocks these use cases:
Want to learn more? We're always here to give you a demo of our product. Email Harry on email@example.com and we'll show you around.
✉️ Monthly newsletter
🤝 Opportunities to contribute
📈 Tactical advice for customer service & CX
📚 Free eBooks, articles and podcasts
💕 +1,000 subscribers