The Art and Science of Conversation Analytics
Why is it still so hard to get insight from customer conversations? We look at the tools and techniques you need to apply and explain why there is so much more to Conversation Analysis than solution suppliers want you to think.
If it were that simple…
One of the frequent conversations I have with clients is about the difficulties they have in extracting anything of value from the conversations contact centre staff have with customers – either over the phone or through Live Chat, for instance. While the complexities of speech transcription and analytics are for another post, the inability to analyse Chat interactions effectively may seem strange. After all, Text Analytics is a well-established field isn’t it?
The clue to the problem lies in the name – ‘Text’ not ‘Conversation’ Analytics. There are plenty of tools that will analyse text for you but not the conversation. Some of these are powered by elements of artificial intelligence – usually machine learning models to assist with Sentiment Analysis or Topic Modelling – but Natural Language Processing (NLP) uses many scientific, linguist modelling approaches. If you want to read in depth, then I suggest you take a look at this work from Stanford University.
What’s the question?
So, Conversation Analytics is what you need to be thinking about, but remember you need the insight not just the analysis. As always, start with the questions you are trying to address and work back to the analysis you need to do and the data you need to achieve that. Here are some of the common questions that we are asked to help with.
- What are customers chatting about? (I include this as it’s the starting point for most clients, but it always leads on to the real, more useful questions below.)
- What are the chat conversations that appear to be resolved but trigger further contact?
- Which advisors are most effective at resolving billing issues and why?
- What should we be looking for in chat conversations to help identify customers at risk of churning?
- Why does our chatbot have better sales conversation rates than some of our live advisors? (Yes, a serious question!)
Notice that answering these relies on sales data, customer account information, customer feedback or telephony system data, not just the information in the text of the conversation. You need a multi-dimensional analysis which many tools simply can’t provide. But let’s look at a couple of ways these tools can help.
Sentiment analysis may help to determine the emotions of the customer as they change through the interaction. That could be combined with Google-like semantic analysis to look for phrases that may indicate a successful resolution.
You can also use semantic search capabilities for key phrases to look for specific topics such as billing errors, product quality issues or payment problems. But it’s important for the tool to track where and for how long topics are discussed, and whether the customer or agent raised them in a conversation. That can influence how relevant they are.
Research has also demonstrated that the biggest challenges in natural language processing come from situations where there is a large vocabulary of words (such as industry sector or product specific language) and bi-polar words which frequently occur in conversations (“Great, now I’ve got to wait another 2 weeks for delivery!”).
Applying machine learning
Machine learning can assist with a number of these questions, but it helps if you are smart with your data. If we think about the repeat contact question above for example, you can hopefully isolate a number of chats that had another contact in the following week, then use agent outcome codes to identify as many as possible that appeared to be resolved in the chat. Assuming you can get a sample of at least a few hundred, you now have the makings of a training dataset for a predictive model. That model can suggest which new interactions may cause repeats. But, it will still be hard to see what aspects of the conversation are triggers for the machine learning to predict a repeat contact. There is little alternative to a systematic manual review and over time, you’ll start to see patterns which can pinpoint root causes. It will be rough, but it will be a start.
So where to start?
You can see these questions can be complicated to answer. The products in the market today have evolved from the need to analyse text snippets in social media and structured customer feedback. They are good at extracting data or classifying content to varying degrees of granularity. You should start with many of the free options and see what you can achieve.
I’ve talked on this previous video post about the different routes you can take to access these capabilities. Your existing Workforce Optimisation or Cloud Contact Centre solution provider may have it as an add on; you can seek out packaged analytics tool providers like Chattermill, MonkeyLearn or Warwick Analytics; or work with self-build tools from likes of Google or Amazon Web Services. You can even build yourself of course, in which case, read the Stanford article. You will need to build a custom model in almost all use cases to cater for your specific vocabulary.
When we work with clients, we recognise that there isn’t one tool that can process all these dimensions and deliver you root causes and clear actions. The element that brings all this very clever technology is the expert conversational analyst. As with many approaches for insight generation, conversation analysis is both a science and an art which needs a human touch – for now.
We help clients to work out the best approach for them, by the way. If you’re just not getting the insight you need, get in touch for…a conversation!
Customers still want to speak to you. How can you deliver great self service experiences on the phone? We are doing this with clients today through Digital Voice. Find out how