/cocoa

Framework for learning dialogue agents in a two-player game setting.

Primary LanguagePython

CoCoA (Collaborative Communicating Agents)

CoCoA is a dialogue framework written in Python, providing tools for data collection through a text-based chat interface and model development in PyTorch (largely based on OpenNMT).

This repo contains code for the following tasks:

  • MutualFriends: two agents, each with a private list of friends with multiple attributes (e.g. school, company), try to find their mutual friends through a conversation.
  • CraigslistBargain: a buyer and a seller negotiate the price of an item for sale on Craigslist.
  • DealOrNoDeal: two agents negotiate to split a group of items with different points among them. The items are books, hats and balls.

Papers:

Note: We have not fully integrated the MutualFriends task with the cocoa package. For now please refer to the mutualfriends branch for the ACL 2017 paper.


Installation

Dependencies: Python 2.7, PyTorch 0.4.1.

NOTE: MutualFriends still depends on Tensorflow 1.2 and uses different leanring modules. See details on the mutualfriends branch.

pip install -r requirements.txt
python setup.py develop

Main concepts/classes

Schema and scenarios

A dialogue is grounded in a scenario. A schema defines the structure of scenarios. For example, a simple scenario that specifies the dialogue topic is

Topic
Artificial Intelligence

and its schema (in JSON) is

{
    "attributes": [
        "value_type": "topic",
        "name": "Topic"
    ]
}

Systems and sessions

A dialogue agent is instantiated in a session which receives and sends messages. A system is used to create multiple sessions (that may run in parallel) of a specific agent type. For example, system = NeuralSystem(model) loads a trained model and system.new_session() is called to create a new session whenever a human user is available to chat.

Events and controllers

A dialogue controller takes two sessions and have them send/receive events until the task is finished or terminated. The most common event is message, which sends some text data. There are also task-related events, such as select in MutualFriends, which sends the selected item.

Examples and datasets

A dialogue is represented as an example which has a scenario, a series of events, and some metadata (e.g. example id). Examples can be read from / write to a JSON file in the following structure:

examples.json
|--[i]
|  |--"uuid": "<uuid>"
|  |--"scenario_uuid": "<uuid>"
|  |--"scenario": "{scenario dict}"
|  |--"agents": {0: "agent type", 1: "agent type"}
|  |--"outcome": {"reward": R}
|  |--"events"
|     |--[j]
|        |--"action": "action"
|        |--"data": "event data"
|        |--"agent": agent_id
|        |--"time": "event sent time"

A dataset reads in training and testing examples from JSON files.

Code organization

CoCoA is designed to be modular so that one can add their own task/modules easily. All tasks depend on the cocoa pacakge. See documentation in the task folder for task-specific details.

Data collection

We provide basic infrastructure (see cocoa.web) to set up a website that pairs two users or a user and a bot to chat in a given scenario.

Generate scenarios

The first step is to create a .json schema file and then (randomly) generate a set of scenarios that the dialogue will be situated in.

The website pairs a user with another user or a bot (if available). A dialogue scenario is displayed and the two agents can chat with each other. Users are then directed to a survey to rate their partners (optional). All dialogue events are logged in a SQL database.

Our server is built by Flask. The backend (cocoa/web/main/backend.py) contains code for pairing, logging, dialogue quality check. The frontend code is in task/web/templates.

To deploy the web server, run

cd <name-of-your-task>;
PYTHONPATH=. python web/chat_app.py --port <port> --config web/app_params.json --schema-path <path-to-schema> --scenarios-path <path-to-scenarios> --output <output-dir>
  • Data and log will be saved in <output-dir>. Important: note that this will delete everything in <output-dir> if it's not empty.
  • --num-scenarios: total number of scenarios to sample from. Each scenario will have num_HITs / num_scenarios chats. You can also specify ratios of number of chats for each system in the config file. Note that the final result will be an approximation of these numbers due to concurrent database calls.

To collect data from Amazon Mechanical Turk (AMT), workers should be directed to the link http://your-url:<port>/?mturk=1. ?mturk=1 makes sure that workers will receive a Mturk code at the end of the task to submit the HIT.

Dump data from the SQL database to a JSON file (see Examples and datasets for the JSON structure).

cd <name-of-your-task>;
PYTHONPATH=. python ../scripts/web/dump_db.py --db <output-dir>/chat_state.db --output <output-dir>/transcripts/transcripts.json --surveys <output-dir>/transcripts/surveys.json --schema <path-to-schema> --scenarios-path <path-to-scenarios> 

Render JSON transcript to HTML:

PYTHONPATH=. python ../scripts/visualize_transcripts.py --dialogue-transcripts <path-to-json-transcript> --html-output <path-to-output-html-file> --css-file ../chat_viewer/css/my.css

Other options for HTML visualization:

  • --survey-transcripts: path to survey.json if survey is enabled during data collection.
  • --survey-only: only visualize dialgoues with submitted surveys.
  • --summary: statistics of the dialogues.

Dialogue agents

To add an agent for a task, you need to implement a system <name-of-your-task>/systems/<agent-name>_system.py and a session <name-of-your-task>/sessions/<agent-name>_session.py.

Model training and testing

See documentation in the under each task (e.g., ./craigslistbargain).

Evaluation

To deploy bots to the web interface, add the "models" field in the website config file, e.g.

"models": {
    "rulebased": {
        "active": true,
        "type": "rulebased",
    }
}

See also set up the web server.