In this episode, we talk with Daniel and Emiel, software engineer and product owner in the customer support domain. In this domain, the focus is to help our customers in the best way possible. But what if we can prevent the customer to feel the need to contact bol.com in the first place, they asked themselves. They realized this can be possible using the analyses of the various customer interactions we have via the Chatbot “Billie”, live chat, phone and email.
For these analyses, they introduced techniques from the Data Science and Machine Learning domain. Natural Language Processing is needed as well as Comprehender Techniques. As a team, they investigated models available in the open source community. In the podcast, we talk about how we adapted them for this purpose followed by the training of these models.
We talk about the four steps to get from idea to the usable information for the product specialists. The product specialist can use the information provided to enhance the product details and descriptions to a level where a minimum of questions from customers is needed.
One of the first deliverables was the introduction of so-called unhappy products report. Products which cause relatively much customer interactions. It presents these products but even more important, possible causes.
Peter Paul van de Beek
Design sprints as part of the way of working to increase the speed and shorten the feedback loop:
Determination of the important words in the text is done by the use of TF/IDF which stands for term frequency-inverse document frequency. It’s part of the Natural Language Processing and to determine the Smart Word Clouds. TF/IDF
BigQuery is used to store the data to be analysed later on.
Predictive Models are being used to improve the shop which should result in a better customer journey with a lower number of customer questions.
Unsupervised Learning is discussed as the way the models are being trained and verified
Clustering Topic Modelling; finding out the latent thems (topics)