Rasa-bot

Table of Contents

• Concepts
• Why intent-based Chatbots
• Examples for intent-based Chatbots
• Discarded Prototype
• Technologies used
• Rasa Components
  • Domain
    • Intents
    • Responses
    • Session Configuration
    • Entities
    • Slots and Forms
    • Actions
  • NLU Data
  • Stories
  • Rules
  • Reminders
• Rasa Architecture
  • Pipeline Components
• Limitations
• Possible Enhancements
• Guidelines
  • Installation with Conda
    • Create new Project
    • Install existing Project
  • Interaction
    • Interact through Shell
    • Interact with REST-Service
    • Interact with Websocket
    • Advanced Remote Server Setup
  • Training
    • Conversation Data Files
    • Interactive Learning
• Sources

Concepts

NLP(Natural Language Processing) helps interact in a human-like way.

We don’t have to match predefined questions character by character. Also it feels more natural for users to write like they would with a human instead of pressing buttons or using forms.

Why intent-based Chatbots?

• They don’t have to follow a specific conversational path. They can adapt to an user changing their mind
• Good at answering specific predefined question-patterns

Examples for intent-based Chatbots:

• Dialogflow1 (originally named Api.ai), developed by Google
• wit.Ai developed by Facebook are used in academic research (Handoyo et al., 2018; Rosruen & Samanchuen, 2019)

Discarded Prototype:

• Custom intent-based chatbot with pytorch and nltk built from scratch.

• Benefits:

  • Lightweight process
  • fast training process

• Drawback:

  • Only single question/answer conversions possible in our prototype
  • Extracting variables from question not possible

Technologies used:

• Rasa

  • with over 25 million downloads, Rasa Open Source is the most popular open source framework for building chat and voice-based AI assistants.

  • Benefits:

    • Open Source Version available
    • Self Host-able
    • Building REST-/ Websocket-Channel
    • Third party system connectors
    • Stories (conversation paths) https://rasa.com/docs/rasa/writing-stories/

  • Drawback:

    • Relative resource intense
    • Long cold start time
    • Training pipeline takes a while.
    • Documentation is lacking

• Linux

  • The rasa chatbot is hosted on a VM with a headless linux distro.

• Ngrok

  • Ssh tunnel for public https connectivity of the local rasa process

Rasa Components

Domain

The domain defines the universe where the rasa-bot performs its function. It defines the intents, the responses and a configuration for conversation sessions and it can also include entities, slots, forms and actions. The domain consists of either one or several YAML files (rasa\domain.yml).

Intents

Intents are categories in which the messages received by the rasa-bot are grouped. They represent the purpose of the user interacting with the bot and are listed in the domain file. It's also possible to use or ignore certain entities for an intent.

Responses

Responses are actions that send a message to a user without running any custom code or returning events. They can be directly defined in the domain file. In order to receive a wider range of responses, it's recommended to add more than one example to each response. Rasa chooses one answer randomly.

Session Configuration

It represents the dialogue between the bot and the user. Conversation sessions can begin when the user starts the conversation, when the user sends a message after a period of inactivity or manually triggered with an intent message.

Entities

Entities are structured pieces of information inside a user message, which have to be extracted by either specifying them or defining regular expressions. By default, entities influence action prediction but it's possible to configure the domain in order to ignore entities for certain intents. Entities roles and groups can also be defined.

Slots and Forms

Slots represent the bot's memory. They could be used to store information provided by the user or coming from the outside (e.g. through APIs). They should be defined with a name, a type and predefined mappings. Slots can also influence the conversation behavior.

Forms are a special type of action intended to help the bot collect information about the user.

Actions

Actions are the things that the bot can actually do. The most common examples are respond to a user, make an external API call or query a database. Responses are included in the domain file, whereas custom actions are only listed on it but they have to be implemented in the actions file (rasa\actions\actions.py).

NLU data

The main objective of NLU (Natural Language Understanding) is to extract structured data from user messages. This information consists of the user's intent and any entities used. It's possible to define some additional information, such as Regular Expressions and Lookup Tables.

NLU data is included in a YAML file, which consists of intents and synonyms (rasa\data\nlu.yml). It should contain at least two examples for each intent to facilitate recognition. The latter can be used to store words which can designate the same entity.

Stories

Stories represent a conversation beetween a user and a bot. They have a name and steps consisting of several intents and actions. When using events, checkpoints and statements they need to be specified. They have to be written to the stories YAML file (rasa\data\stories.yml).

Rules

Rules are used to handle small specific conversation patterns, but they cannot generalize to unseen conversation paths, unlike stories. Their overuse is not recommended and the RulePolicy should be included in the model configuration. They are included in the rules YAML file (rasa\data\rules.yml).

Reminders

Reminders are used to reach out to the used after a set amount of time in order. A reminder has to be scheduled and a reaction to it needs to be specified. It's also possible to define an action in order to cancel its effect.

In a previous phase of the project reminders were implemented to trigger a specific bot response at a given time to simulate that the bot had taken the lead to change the subject or talk about something in particular. This option was disabled since it does not suit the specifications of the project.

There are other components, some of which have been already mentioned, which will not be explained further since they have not been used in this project.

Rasa Architecture

Rasa architecture is scalable. The two main components are Natural Language Understanding (NLU) and dialogue management, respectively represented in the following diagram as NLU Pipeline and Dialogue Policies. The first is responsible for handling intent classificaton, entity extraction and response retrieval; and the latter decides the next action in a conversation based on the context.

Diagram Rasa architecture overview

Pipeline Components

• Language Model. The components MitieNLP and SpacyNLP load pre-trained models which are necessary to use pre-trained word vectors in the pipeline.
• Tokenizers split text into tokens. Tokenizers that may be used are WhitespaceTokenizer, JiebaTokenizer - for Chinese language -, MitieTokenizer, SpacyTokenizer.
• Featurizers transform raw input data into a feature vector and are divided into sparse featurizers and dense featurizers. Featurizers that can be used are MitieFeaturizer, SpacyFeaturizer, ConveRTFeaturizer, LanguageModelFeaturizer, RegexFeaturizer, CountVectorsFeaturizer and LexicalSyntacticFeaturizer.
• Intent Classifiers are responsible for assigning one of the intents defined in the domain file to messages from the user. Some of them are MitieIntentClassifier, LogisticRegressionClassifier, SklearnIntentClassifier, KeywordIntentClassifier, DIETClassifier and FallbackClassifier.
• Entity Extractors are responsible for extracting entities from incoming user messages. Some entity extractors are MitieEntityExtractor, SpacyEntityExtractor, CRFEntityExtractor, DucklingEntityExtractor, DIETClassifier, RegexEntityExtractor and EntitySynonymMapper.
• Selectors predict a bot response from a set of possible responses.
• Custom Components. It's also possible to create and add a custom component to the pipeline.

Limitations

• A rasa-bot uses pre-trained NLU data, which means that only entities declared in the NLU data can be recognized and assigned to incomming user messages.
• Rasa is not compatible with Python versions above 3.8v.
• Rasa's universe is wide but relatively poorly documented, so that it requires time to learn about all of its options.

Possible Enhancements

• Add corpus
  • Example: Building a rasa chatbot to perform a movie search
• Host webui and text to speech

Guidelines

Installation with Conda

1. Open a terminal window and run conda create --name <name> python=3.8 to create a conda environment.
2. Run conda activate <name> to activate the created environment.
3. Run pip install rasa to install rasa.

Create new Project

4. Create folder rasa and change to this directory.
5. Run rasa init to create a new project with example training data, actions, and config files.

Install existing Project

4. Change to rasa directory.
5. Run rasa train to train a new model according to the given data.

Interaction

Interact through Shell

1. Open two terminal windows and activate the environment in which rasa has been installed.
2. Run rasa run actions in one of them to use registered actions.
3. Run rasa shell in the other one to load the trained model and interact with the bot on the command line. To stop the process run /stop.

Interact with REST-Service

1. Start server with (only) REST interface: rasa run --connector rest ( bind specific local ip: -i 192.168.50.150)
2. Start action server: rasa run actions
3. Interact with REST Interface: http://0.0.0.0:5005/webhooks/rest/webhook

• Send a POST Message to the REST Server with a JSOn Body:
  • { "sender": "test_user", "message": "Hi there!"}
  • Response: [{"recipient_id": "test_user", "text": "Hey! How are you?"}]

Interact with Websocket

1. Start server with CORS enabled: rasa run --enable-api --cors "*"
2. Start action server: rasa run actions
3. Start a webserver of your choice inside the subfolder ./webclient

Advanced Remote Server Setup

Instructions found in this document: https://github.com/tsrodf/rasa-bot/blob/main/server/server_setup.md

Training

Conversation Data Files

1. Add new intents, utter_responses, stories, rules and/or actions to the proper files.
2. Open a terminal window and activate the environment in which rasa has been installed.
3. Run rasa train to train a model using provided NLU data and stories. Models are saved to ./models.. This should be done every time the NLU data is modified.
4. You can run rasa data validate to check if the given NLU data is correct. This function works faster than rasa train, but it's necessary to train a model in order to interact with the bot.

Interactive Learning

1. Open two terminal windows and activate the environment in which rasa has been installed.
2. Run rasa interactive in one of them to to activate the interactive learning tool.
3. Run rasa run actions in the other one to use registered actions.
4. Interact with the CLI in order to train the bot with new NLU data.

Sources

• Luo, Bei et al. (2021) A critical review of state-of-the-art chatbot designs and applications
• Rasa Documentation