TacoBot

This is the codebase for TacoBot. The code structure in this repository is inspired by the implementation of Chirpy Cardinal. We thank the authors for their valuable work.

Getting Started

• Check the code out, start here
• Check the datasets we collected, start here

Code Structure

Under the ./code folder:

agent: When you run tacobot, you will create an agent. Agents manage data storage, logging, user message input, bot message output, connections to remote modules, and calls to the handler. The agent is provided:

• local_agent.py: an agent that stores data and runs remote modules locally.

servers: Contains the code needed to run tacobot servers

• servers/local/shell_chat.py: script to build docker modules locally and run chat in a loop.
• servers/local/local_callable_manager.py defines the LocalCallableManager class, which is used to run docker containers locally
• servers/local/local_callable_config.json defines the ports, dockerfiles, and urls associated with each container

taco: This directory contains the bot’s response generators, remote modules, and dialog management. The core logic of the bot is here. Code in this directory is invariant of agent specifications.

taco/annotators When a user utterance is input, all annotators are run on it and their results are stored in state, so that they can be used by the response generators. Annotations include dialog act and user emotion, among others.

taco/core The bot’s core logic components. Highlighted files are:

• dialog_manager.py: this contains the function get_response_and_prompt, which runs all response generators, ranks their responses, and returns the highest ranking response and prompt, and the function execute_turn which loads the rg states from the previous turn, updates the state based on the response and prompt chosen by get_response_and_prompt and then returns the bot’s next utterance
• handler.py deserializes the state, runs the NLP pipeline, updates the state based on it, calls dialog manager’s execute_turn, and then serializes the state
• taco_ranking_strategy.py Logic for ranking responses and prompts
• state.py: The State class defines what should be stored in each state and contains functions for serializing/deserializing the state.
• user_attributes.py: The UserAttributes class defines which user attributes should be recorded and contains functions for serializing/deserializing user attributes.

taco/response_generators: Contains all response generators used by the bot. More detail can be found in the Creating a Response Generator section

Creating an Agent

Agents manage the bot’s data storage, logging, message input/output, and connections to remote modules. The agent class provided, local_agent.py stores data locally and inputs/outputs messages as text. By defining your own agent, you can alter any of these components, for example storing data in a Redis instance, or inputting messages as audio.

Highlighted features of the LocalAgent are: init function, which initializes

• last_state and current_state dicts These are serialized/deserialized by the functions in taco/core/state.py. If you change their attributes in your agent, then you should also update state.py
• user_attributes dict, which contains
  • user_id: unique identifier for the user
  • session_id: unique identifier for the current session
  • user_timezone: the user’s timezone (if available) which is used by response generators to create time-specific responses, e.g. “good morning!”

persist function

• Manages storage of the state and user_attributes. If you want to store things non-locally, you would make this change here should_launch function
• Determine whether to launch the bot, for example based on specific commands should_end_session function
• Determine whether to end the conversation, which may also be based on specific commands or heuristics process_utterance function
• Retrieve the current state, previous state, and user attributes from your storage
• Call handler.execute() on the current state, previous state, and user attributes, which returns updated states and a response
• Persist the updated states in your storage
• Return the response and current state

Creating a new Response Generator

To create a new response generator, you will need to

1. Define a new class for your response generator
2. Add your response generator to the handler

Defining a Response Generator class

You will need to create a new class for your response generator. To do this,

1. Create a folder which contains new_response_generator.py in taco/response_generators which defines a NewResponseGenerator class
2. Set the class’s name attribute to be 'NEW_NAME’

Adding a Response Generator to the Handler

In order for your response generator to be called, it needs to be added to a) your handler and b) the response priority list. To do this,

1. Add MyNewResponseGenerator to your handler’s list response_generator_classes in your agent. If you’re using the local agent, you would add this to local_agent.py
2. Using the name you declared in your response generator class, set the following in response_priority.py:

Datasets

datasets: In this folder, we open-source the datasets we collected during the process of tacobot development, which covers multiple components, including natural language understanding, question answering, etc.

Domain Classifier

Inside the folder datasets/domain-classifier, you will find the dataset for domain classifier. Each line in the data file represents a unique record and is structured as:

{
    "utterance": "how do i paint a wall",
    "domain": "diy"
}

Key Descriptions:

• utterance: Represents the text of the user utterance.
• domain: The domain tag for the given utterance, either 'cooking' or 'diy'.

Intent Classifier

Inside the folder datasets/intent-classifier, you will find the dataset for intent classifier. Each line in the data file is structured as:

{
    "utterance": "do you like a movie",
    "intents": ["QuestionIntent"]
}

Key Descriptions:

• utterance: Represents the text of the user utterance.
• intents: A list containing the intent labels corresponding to the given utterance.

Question Type Classifier

Inside the folder datasets/question-classifier, you will find the dataset for question type classifier. Each line in the data file is structured as:

{
    "sentence": "which ingredient can i use as an alternative for bread",
    "label": "SubstituteQuestion"
}

Key Descriptions:

• sentence: Represents the user's question text, including any accompanying context.
• label: The tag of the question type.

Machine Reading Comprehension (MRC)

Inside the folder datasets/mrc, you will find the dataset for MRC QA. Each line in the data file is structured as:

{
    "url": "https://www.wikihow.com/Make-a-Water-Filter",
    "title": "How to Make a Water Filter",
    "step_count": 4,
    "step_context": "Put the coffee filter over the mouth of the bottle and tighten the cap over it ...",
    "question": "How to keep the bottle steady?",
    "answer": "Put the bottle cap-side-down into a mug or cup. This will help keep the bottle steady while you fill it."
}

Key Descriptions:

• url: The url of the WikiHow article.
• title: The title of the corresponding WikiHow task.
• step_count: The index of the step number in the task.
• step_context: The step content in the task as the context.
• question: Represents the user's question text.
• answer: The answer of the question.

Citation

@article{mo2023roll,
  title={Roll Up Your Sleeves: Working with a Collaborative and Engaging Task-Oriented Dialogue System},
  author={Mo, Lingbo and Chen, Shijie and Chen, Ziru and Deng, Xiang and Lewis, Ashley and Singh, Sunit and Stevens, Samuel and Tai, Chang-You and Wang, Zhen and Yue, Xiang and others},
  journal={arXiv preprint arXiv:2307.16081},
  year={2023}
}