Chinese conversational assistant robot
RASA may be the best open source to help people build their own auotomatic assitant robot. Here, I will show you how to build a very basic chat robot that handle multiple tasks with rasa.
- Conversational, not just dumb **asked-and-reply** pattern, but can follow the conversation in a very natural way. - Can extract the certain information pieces by using slotting filling and name entity recognition. - Handle multiple missions and tasks. Features
Frist, clone the reposity and download dependeices: How to
pip install -r requirements.txt
Then, train your model
Start the actions server:
rasa run actions
Ok, open another terminal, just input
Here you go(cuz here we call a outside api service to do some NER for us, you may encounter a error, which will be explained below).
Before we start, make sure you have already knew the basics of [RASA]( Under The Hoodhttps://rasa.com/docs/).
Be conversational with story
Traidiotional robot follow a asked-and-reply pattern, which robot can only responde after a user input, and most of time , the response given by robot allways has nothing to do with the previous talks, which make the experience of talking with a robot like talking with a stupid robot. Whereas, Rasa can maintain a converational state by using its memery policy and can reponde like acting a Shakespeare play. For eamples, when the robot find you are sad(what we call
intent recognition, nothing more than a text classification), the robot will throw out a joke and cheer you up, then, it will ask your feedbacks, if that helps, the robot will send a poistive messasage, otherwise it will do something else. To make that happen, Rasa use a so-called
story mechanism to make the conversation more conversational.
story can be written as following(check the
stories.md in ):
## sad path 1 * greet - utter_greet * mood_unhappy - utter_cheer_up - utter_did_that_help * affirm - utter_happy
sad scenario has three certain stages to interacte with our users, fell free to and more stages.
Retrieve certain informations
Often , the robot need to extract some imformation pieces to complete a certain mission, let's say you wanna call a taxi, then the robot need to know where you wanna go. There're two ways to achieve that:
- Slotting filling by asking user some questions until all slots being filled.
- Also ask the user some questions, but call a NER api to extract informations. Here we use the scond method. Rasa allow us to write custom NLU Components, here I called a bert-ner api service to extract neccessary informations for me, which is locations here. You can check this repositry to easily train and build your own ner models here
After knowing the intent of our user, and got neccessay information, the robot then can actually do some staff for us. But here, for simplicity, I just let the robot responde with some messages, but you can write your own actions.