Some companies choose to rely on the help of a tool such as QBox. QBox is another way of testing your chatbot to help improve your its accuracy and performance — in a matter of minutes. It analyses and benchmarks your chatbot training data and gives insight by visualising and understanding where your chatbot does and doesn’t perform, and why. You can see your chatbot’s performance at model level, intent level and utterance level, even word-by-word analysis, in clear and easy-to-understand visuals.
For example, we built a very simple banking bot based on common customer-banking questions. But I noticed that some user questions requesting a contact telephone number didn’t always return the correct intent with good confidence — despite having training data to that effect. I was able to diagnose the problem quickly by doing a deep dive into the actual training data and getting word-by-word analysis, using the visuals that QBox provides:
This helped me to diagnose why the training data wasn’t working as expected.
From the colour-coded words (on the left), I had a good understanding of the word density, and could see I had several instances of “telephone” in a different intent, and a lack of variations on asking for a telephone number in my contact_us intent.
You can get an even deeper word-by-word analysis by using the QBox Explain feature (bottom illustration). This chart shows each word according to its influence on the prediction. In this illustration, you can easily see the biggest influencer to return an incorrect intent is “telephone”, and the lack of influence this word has on my Contact intent. This information clearly showed what the problem was and what I needed to do to fix it: add a little more training data to my Contact intent to strengthen the concept of asking for a telephone number.
Once you make changes to your chatbot, you can test your model again through QBox to validate those changes and see what effect they’ve had on your model overall — whether good or bad.
To find out more, visit QBox.ai.
Nobody would dispute the importance of ensuring your chatbot performs well to secure customer satisfaction. More recently, we’re seeing that it’s becoming a business requirement to understand the NLP model performance, analysing regression, ensuring clear KPIs are set and building your model development as part of a DevOps (MLOps) strategy. Whichever technique you use, it should help you on the path to improving performance, and ultimately your confidence in your model.