Email bot training
Pega Email Bot uses natural language processing (NLP) to analyze and learn from patterns of conversation between customer service representatives (CSRs) and customers. Pega Email Bot learns from repeated behaviors — the more emails it processes, the more accurate it becomes in helping a CSR respond to customers.
For each topic added to the channel, you must train Pega Email Bot to interpret emails. By providing examples to Pega Email Bot and manually classifying them with a topic, you train the NLP model to understand these topics and classify similar emails.
In Pega Customer Service™, both the CSR and the customer service administrator (CS admin) contribute to the training of the email bot.
A data scientist can also train Pega Email Bot by working directly with NLP models in Prediction Studio.
CSRs work in the Interaction Portal. CSRs train the model in real-time by triaging and mapping content. They help Pega Email Bot learn by categorizing unmapped fields and correcting mappings that are incorrect.
For example, suppose Pega Email Bot fails to identify the customer’s street address. The CSR can make the correction when working with the Account Address Change case.
When a CSR adjusts an entity mapping, the change is saved as a training record in the email channel.
Training record collection and review
The CS administrator works in the App Studio. The admin can collect, curate, and triage training records in the email channel.
A training record contains the email content and the NLP analysis, including the mapped entities and the associated service case.
If a CSR adjusts the entity mapping when processing a case, a training record is created and identified as a triaged record. The admin can review the record and make any further changes.
A training record can also be added manually by the administrator. As each training record acts as an inbound email, adding records manually lets you train the email bot in real-time.
When an email bot is first instantiated, the CS administrator adds training data to the channel by adding or importing records. The admin can also classify each record by associating it with the appropriate service case.
Once the admin classifies the training data, the data is marked as reviewed and added to a queue. The NLP model is then rebuilt to include the all-new reviewed records, as well as any existing training data that was already present within the model. Each time the model is rebuilt, a new score is calculated, which represents the model's overall accuracy across all topics..
Admins can monitor a model for accuracy, rebuild a model, and export or import the models across different Pega environments.