HOW IT WORKS
Commercial Natural Language Understanding libraries, such as Dialogflow and Alexa, are used to develop conversational agents in our projects.
Children will be provided with two opportunities (see green arrows in diagram) to answer one set of questions, with the follow-up prompt rephrasing the original prompt into a multiple-choice format. The scaffolding mechanisms combined with edge case handling ensures the system would still be able to drive the conversation along the context even if the CA does not understand the exact utterance of a participant.
This language training model is optimized constantly during the course of field testing as we collect more data on children’s responses. Specifically, we modify the intents (e.g., add more intents to encompass other common response categories) and include more training phrases to increase the accuracy of intent classification.
- The agent learns to understand children’s responses both from the pre-trained language models already built into the commercial libraries as well as training phrases that we provide, which are sample phrases of what children may say to respond to a particular conversational prompt.
- For each conversational prompt, we predefine categories, or intents, that we want the agent to classify the utterances into.
- After the agent classifies the child’s responses into one of the intent categories, differentiated feedback is given based on the classification.
Children will be provided with two opportunities (see green arrows in diagram) to answer one set of questions, with the follow-up prompt rephrasing the original prompt into a multiple-choice format. The scaffolding mechanisms combined with edge case handling ensures the system would still be able to drive the conversation along the context even if the CA does not understand the exact utterance of a participant.
This language training model is optimized constantly during the course of field testing as we collect more data on children’s responses. Specifically, we modify the intents (e.g., add more intents to encompass other common response categories) and include more training phrases to increase the accuracy of intent classification.