Using Advanced Dialog Features in Watson Conversation
Bots are smarter than ever, but sometimes they’re not able to show off how smart they really are. Humans say and type in unpredictable ways, that even an expert conversationalist can’t always account for. What if a user asks a bot two things at once? Even a human would have to process one request at a time. What if a user is going through one dialog flow, and decides they want to do something else instead? A human would have to stop the process they’re going through and switch gears. Designing a dialog is no easy feat. Luckily there are tricks and tips that make sure your bot sounds as smart as it really is. This post will show you how using the Watson Conversation service (which you can try out for free).
Watch this step-by-step video on how to utilize the more advanced features of Conversation or continue reading through this post.
This tutorial goes over how to handle two intents, manage handlers (how your bot will respond when the user’s answer to a prompt is not found), use pattern entities, and add a counter. You can find the workspace json for this tutorial here to follow along.
Handling two intents
If a user asks a question or a request that contains two intents, for example, placing and order and making a return, the bot (and humans) will need to take it one at a time. To do this, you want your bot to be able to detect if multiple things are being said, and then clarify which one to process first.
In your dialog tree, create a node to disambiguate between the intents. This node is checking to see if the top two intents returned are higher than 40% confidence (40% was chosen because there is not much training data). If yes, then the bot can respond with “Sorry, I can only handle one request at a time. What would you like to do first?”
Manage handlers
During a conversation, a user may change their mind and switch topics. For example, a user may be placing an order but then decide to return something first instead. With slots, a feature to gather information in a single node, the bot looks for the specific answers. If the specific answers are not found, your bot can look for handlers, or global intents within the node, and respond accordingly. So if a user decides to cancel the order half way through the flow, and a handler is set for a #cancel intent, the bot can stop collecting the necessary information and exit the node.
To add a handler, go to your node, and click ‘Manage handlers’. Here, you can add the #cancel intent and the appropriate response.
Then, click the settings icon, and make sure the bot is skipping to the response. This ensures that the bot will exit the node, instead of collecting more variables.
Finally, you also want to ensure that any variables collected have been reset. This means, if your user was in the process of ordering an Apple phone, the bot has saved those variables `$brand=Apple , $device=phone` and will not collect that information next time, unless otherwise stated. Create a ‘Jump to’ child node that’s set to “true” (so this node is always fired after the #order node).
In our example, we’ve set a confirmation variable, so the bot can confirm with the user at the end of a flow. Open the json editor and set the context variables to “null”.
Pattern entities
In several use cases, such as this bot, some answers require alpha-numeric text or symbols (i.e. emails, order numbers, etc). Instead of training the values for every possible combination, pattern entities allows you to set the pattern in regular expression and then able to detect that value. In our example, we created a pattern for an order number and email.
Go to the entities section of the tooling, enter the value and choose ‘Patterns’.
If you’d like to store a pattern entity as a context variable, you want to save the value found rather than the expression. To do so, just add “.literal” syntax after the variable, for example, “@email.literal”
Counter
Many times, a user will ask or say something completely off topic or the bot simply does not have the capability to understand. Rather than the bot repeatedly respond with the same responses like “I don’t understand” or “Can you rephrase,” you can have your bot know when it’s the right time to pass off to a live agent or end the conversation. Adding a counter tells the bot that if it doesn’t understand the user after a certain amount of times, escalate to an agent or end the conversation.
In dialog, create a context variable in the node you want to count from. In the example, we added the “agentCounter” (can be named anything) variable in the Anything else node.
Add a node above (we named it “escalate to agent”), that recognizes if the context variable $agentCounter is above 2, respond accordingly. This number can be set to whatever you’d like.
Finally, we then need to create the counter and set the variable to 0 in the ‘Welcome’ node.
While this post reviews how to use these features, make sure to watch the video for more details.
