Builder allows a non-technical expert to easily build and train chatbots which can accurately respond to a wide variety of user inquiries by utilizing various Watson API's. Builder will guide you through each of these steps as you build your chatbot!
How does Builder Learn?
Builder learns through matching training questions with Responses, using Natural Language Understanding. A user does not need to ask the question in the same way as the training questions in order for Watson to deliver a high-confidence match.
What types of Training Questions should I include?
Should mimic a variety of phrases that an end user might ask about a subject within a conversational experience.
Alternative questions are used as additional training material for Watson. Include industry-specific vocabulary to augment Watson’s knowledge.
WATSON-GENERATED ALT QUESTIONS- BETA
Watson will automatically generate additional questions for your conversation via the Scrape Tool, using any Q&A formatted content it finds from a web page. You can then choose to add these questions to the training material or discard them.
On the publish tab, run an Accuracy Test to check the viability of your conversation. When the accuracy reaches 75% or higher, you can submit your conversation for publishing.
How does Builder Respond?
Builder surfaces a Response from your Corpus based on a confidence threshold.
What types of Responses should I include?
Builder surfaces a Response when a user statement/input aligns directly with a Response within the Corpus. Builder has a high level of confidence in the matching Response.
LOW CONFIDENCE RESPONSE
Builder surfaces a Low Confidence Response when Builder has a low level of confidence in a match between a user statement/input and a Response within the Corpus. Builder weighs related Responses and surfaces a Redirect query to clarify the user’s statement/input, along with the Tags of possible Response matches. This helps the user find the information they are seeking. (Find and edit within Advanced Edit Mode).
Builder does not require any additional training on the infinite possibilities that users could input which are off-topic for the conversation, and thus do not match with a Response in the Corpus. Out of Scope Responses are rotated at random . Their text should acknowledge the user has gone off topic and steer them back on track within the conversation. (Find and edit within Advanced Edit Mode).