Context: it is important for CC analysts to quickly understand what topics are included in the dialog corpus in order to quickly automate their work. Manually dividing such cases into topics is a resource-intensive task.
Decision: a system of hierarchical clustering of contact center dialogs is proposed, which allows creating a taxonomy of intents
Results:
Integration into the BI Department for automation of customer contact centers:
Reducing the load on the analyst to allocate automation classes by 80%.
Harmonization of taxonomy:
Reducing the time to identify a new category by 60%;
Allocation of new intents in the flow of requests with a quality of 70%.
The selection of entities and the analysis of interpretability:
Selecting named entities and filling in the client card is 10% more accurate;
Increase in the share of interpreted intents relative to the old model by 40%.
To build the model, we used:
Dialogs for multiple contact centers;
Assessor markup of paraphrases for each dataset.
Simulation result:
Model for soft hierarchical clustering of dialogs;
Final taxonomy of categories with a description;
A server-side custom application with an API interface.
Customer: Contact Center, Telecom
Technology stack: TopicNet, BigARTM, Flask, Python, PyTorch.