Available in Classic and VPC
Click the [Advanced settings] button at the top of the conversation information to set various detailed features and build a variety of conversations. You can create multi-turn conversations by setting contest or change the conversation type to a general conversation or a task. Alternatively, you can use the negative settings to compose negative data that helps the learning model, or set separate feedback answer messages for each conversation.
The following items can be set in the advanced settings window.
- Context settings
- Conversation type settings
- Conversation group settings
- Task flow settings
- Negative settings
- Feedback response settings
If you've created welcome messages, multiple general conversations, tasks, failure messages, etc., for building a chatbot, then you need to connect these conversations. This is called context settings. Context is a link that connects between conversations, and is used for understanding the context of connected conversations. Enter the same context in the output and the input of the conversations you want to connect with.
If the output context of Conversation A and the input context of Conversation B match, then after Conversation A is matched, Conversation B, which has a matching context, takes precedence over other conversations without context in matching. If there is a conversation connected through multiple choice answers, then the conversation connected to the multiple choice button will be matched as a priority.
The following shows how to set context.
- Enter the conversation information by referring to Register conversation.
- The advanced settings button will be enabled only after the conversation information is registered.
- Click the [Advanced settings] button from the conversation.
- Register context information in the [Context settings] tab.
|Context strength settings||Context strength settings
* Hard: Prevent leaving the set context flow
* Soft: Allow leaving the set context flow
|Input context||Enter input context.
A conversation model determines the conversation context by comparing the entered input context and the context left in the user's log from previous conversations. If the context left in the user log matches the input context of the conversation, then the model determines that the context is connected.
|Output context||Enter output context.
As the context to be left in the user log once the conversation is finished, it will be deleted after the set number of counts or time.
* Action: Set how to process the output context.
- Add: Add the context to the user log.
- Maintain: Maintain the context count that should have been deducted in this turn without deduction.
- Delete: Delete the context from the user log.
* Count limit: Limits the lifespan of the context to be left in the user log by count. The context count is deducted with each turn passing, and when all available context counts are deducted, the context is removed from the user log. (However, asking back in a slot conversation does not affect the lifespan of the context.)
* Time limit: Limits the lifespan of the context to be left in the user log by time. Once the set time expires, the context is deleted from the user log.
While context is maintained, the slot information collected is remembered, so it is possible to create a chatbot service that can respond to more complex conversations by using the slot information collected from previous conversations. For more details on how to preserve slot information using context, refer to Using task.
For instance, let's say there are four conversations as below.
- Conversation A: Task to make a reservation. The conversation checks for the date, number of people, and time, and confirms the reservation with the final answer
- Conversation B: The conversation where the chatbot returns the confirmation of the reservation message as the answer when the user enters "Yes"
- Conversation C: The conversation where the chatbot returns the cancelation of the reservation message as the answer when the user enters "Yes"
- Conversation C: The conversation where the chatbot returns the message "I see" as the answer when the user enters "Yes"
After receiving the reservation information through the task to make a reservation, the chatbot runs Conversation A in which it asks whether to proceed with the conversation. If the user answers with "Yes," then the chatbot will search for a conversation with "Yes" entered in a question. If the conversations haven't been connected as contexts, then it will return a random answer among the three conversations with the "Yes" question. However, if you connect Conversation A and B with the context "Make reservation," then entering "Yes" as the answer for Conversation A will lead the user to Conversation B.
Depending on the chatbot's service scenario, you can choose a general conversation or task.
- General conversation: When creating a simple conversation where chatbot responds using an registered answer by determining the user's intent from an utterance, general conversation is appropriate. Most chatbots can only use general conversations to create a service.
- Task: Select task if you want to create a conversation with a complex flow that collects slots through interactive conversations, and performs certain tasks based on the slots collected. For more information about how to build a task, refer to Using task.
The conversation type can't be changed after the answers are registered. The conversation type must be changed before registering answers.
The conversation group settings tab is used to display conversation flow in a separate page space on the conversation canvas. Thus, grouping the flow into smaller groups is recommended. The "Home" conversation group is the default value unless a separate conversation group is specified.
You can set the flow between tasks and general conversations in the task flow settings tab. For more information on scenarios using task flow, refer to Using task.
It's an option enabled in task conversations. If the user asks a question unrelated to the slot during a slot filling process of a task, then it's allowed to temporarily leave the task and find an appropriate answer in general conversations. However, answers with multiple choice forms or open-ended forms registered can't be matched to other tasks.
It's an option enabled in general conversations. If the user asks a question unrelated to the slot during a slot filling process of a task, then it's allowed to temporarily leave the task and interrupt the task with a general conversation related to the user's question.
Return to task
It's an option enabled in general conversations. It is configured so that after leaving the task and responding with an answer from a general conversation, the chatbot returns to the task.
Negative settings are used to train a conversation model to learn the subtle differences between similar conversations. Select a conversation in which similar questions are registered, or a different conversation that the model can't distinguish, and register it as a negative. This allows the model to differentiate between the registered negative conversation and the current conversation. Up to 20 conversations can be registered as negatives.
You can select a feedback response message to survey the customer satisfaction for the conversation in the feedback response settings tab. If nothing has been set here, then the feedback set as the default response message will be returned. For how to add a feedback response message, refer to Register feedback.