This is the example bot to demonstrate Rasa's support for custom components--specifically, custom intent classifiers.. Its sample code from the post "How do Chatbots Understand?", which can be found on Medium.com.
It builds upon our existing example, which can be found here.
- addition of fallback mechanisms
- addition of simple, single stage and two stage fallback mechanisms
- support for asking the user to rephrase themselves
- support for asking the user to select from some suggestions
- custom intent classifier
Logistic Regression classifier
👨 : hey (should go to intent "greet")
prediction:
-----------
Received user message 'hey' with intent '{'name': 'greet', 'confidence': 0.9377497415166074}' and entities '[]'
🤖 : Hello! Could you please provide your contact information?
This project follows the format of a standard Rasa project. There's a directory called data for training data like nlu, stories, and rules.
There's a directory called actions, which contains all your custom actions.
You'll also find the domain.yml file, which mentions all your intents, entities, slots, responses and actions.
Finally, there's the config.yml file, which specifies the components your bot is comprised of.
The config file has a modification compared to the previous versions of this bot: a custom component. It'll look like this now:
..
..
- name: DIETClassifier
entity_recognition: true
intent_classification: false
epochs: 5
- name: lr_classifier.LRIntentClassifier
max_iter: 100
..
..
You can see that we've also added a intent_classification: false field under the DIETClassifier, meaning that we'll be using it only for entity extraction, since our custom componenent will be doing the intent classification.
The code for the custom intent classifier which is based on the logistic regression classifier is located in lr_classifier.py.
Lastly, we look at the results folder. It contains an evaluation of the performance of our bot's intent classification and entity extraction. There aren't any results for the Core part of our bot since we don't have any test cases for that.
- Clone this repo
- Navigate to the RasaChatbot directory
- Install rasa>=2.6.2 in an env.
Modify the files in data/ or the domain.yml file to play around.
Before training the bot, a good practice is to check for any inconsistencies in the stories and rules, though in a project this simple, it's unlikely to occur.
$ rasa data validate
To train the bot, we simply use the rasa train command. We'll provide a name to the model for better organization, but it's not necessary.
$ rasa train --fixed-model-name logistic_regression_classifier_model
To test your bot, open a new terminal window and start a rasa shell session.
$ rasa shell
This will let you chat with your bot in your terminal. If you want a more interactive UI and a little more debugging information like what intents were identified and what entities were extracted, you can use Rasa X.
The code for the custom intent classifier which is based on the logistic regression classifier is located in lr_classifier.py. You can play around with the different parameters by modifying the fields in config.yml. You can also plug in your own sklearn models directly. Simply change the model used in the train method in lr_classifier.py to, say, MultinomialNB to try the Naive Bayes sklearn classifier.
After this, retrain the bot to test it. As always, modify the files in data to add more intents or more data to existing intents. Keep in mind that this is a custom intent classifier, so it won't extract entities nor will it affect the response the bot gives, directly.
Rasa provides support for evaluating both the NLU and the Core of your bot. All we have to do is create some test data and run rasa test . To run the tests, execute:
rasa test
The results folder will be updated.
You can find me on medium here.