One of our previous chatbot explainer posts discussed that intent detection–the chatbot’s ability to recognize the correct intents from the customer’s inputs–is the most fundamental functionality for chatbots. Intent detection is usually enabled by a machine learning model which is built by the enterprise user through the creation of intents and a large body of utterances for each intent. However this process can be greatly simplified and accelerated by using a zero-shot or few-shot model. These models are classified in the Training category of chatbot product features.

The enterprise user can use a zero-shot or few-shot model for the machine learning engine of their chatbots by integrating with a proprietary or third-party LLM provider such as OpenAI or Anthropic. For the zero-shot model, the enterprise user only needs to provide intent names but not any utterances, thus saving the user time and effort in creating them. The model enables the chatbot to identify the correct intent from the customer’s input by its semantic similarity to the intent name alone. For example, the enterprise user only needs to create an intent with the name “Dispute credit card transaction,” and relevant customer inputs such as “I see transactions on my bill which I didn’t pay” and “How do I dispute invalid items on my bill?” will be matched to that intent.

While the zero-shot model can accelerate the chatbot development process, one downside is that if the customer input fails to match to the correct intent given the limited training, there is no way to train it to do otherwise. The few-shot model, on the other hand, provides a solution to improve the accuracy of intent detection by requiring only a few utterances for the intents. Using the few-shot model, the enterprise user can start by only providing intent names, and after testing the model they can add the failed inputs as utterances for their associated intents. For example, for the intent “Dispute credit card transaction,” the customer input “I see transactions on my bill which I didn’t pay” is correctly matched to it while the input “Why is this charge on my bill? I don’t recognize it” fails to be matched to it. The user can add the failed input as an utterance for that intent in order to train the model.

The screenshot above shows the feature to select zero-shot and few-shot models from an example chatbot vendor – Kore.ai.

Our chatbot explainer series shares our knowledge of chatbot product features. If you have any questions, please reach out to me directly and follow me to get notifications on future posts.

Dong Liu, Founding Analyst
dong@daybreakinsights.com

@dongliu1226