I am currently helping our bot team use the ‘Intent Miner’ to identify new customer intents from our historical chat transcripts. I am seeing a lot of ‘Coherence’ issues where the miner is grouping unrelated utterances into the same intent. This makes the intents almost impossible to train. How can I improve the quality of the intent mining results? Is there a way to filter out ‘Noise’ words or provide ‘Anchor’ phrases to help the AI understand our specific technical domain?
Hello Pri66. I am a GC trainer and I deal with these bot training issues every day. The Intent Miner is a ‘Unsupervised’ learning tool, which means it is doing its best without knowing your business context. You should definitely use the ‘Stop Words’ feature to filter out common phrases like ‘Hello’ or ‘Thank you’. Also, instead of mining your entire history, try mining small batches of conversations that were routed to specific technical queues. This provides a natural ‘Anchor’ because you already know the general topic of those calls.
I have seen these coherence issues impact our topic detection as well. To follow up on Isa53, you should also look at the ‘Clarity’ score for each mined intent. If the score is low, do not bother training it. Instead, take one or two utterances from that group and manually create a new intent. Then, use the ‘Intent Health’ tool to find more similar phrases. It is much more effective than trying to fix a ‘Messy’ mined intent!