Task-Oriented Dialogue with In-Context Learning

Code for the Paper Task-Oriented Dialogue with In-Context learning.

Results are stored in line-counts.json and the test-results dir. To render a table of results, pip install prettytable, and run:

make show-results

You can also see the test conversations in the e2e_tests dir.

Installation & Setup

You need a Rasa Pro license to run the code. You can get the free Rasa Pro Developer Edition here.

Install poetry, and then run

make install

Set the env var RASA_PRO_LICENSE=<your rasa pro license key> If you want to run version_1, you'll need to set the env var OPENAI_API_KEY=<your openai api key>. Alternatively, you can edit version_1/config.yml to use a different LLM. Follow the instructions here.

Some of the tests require duckling to pass. To set this up, run a duckling server, and set the url in version_1/actions/check_restaurant_availability.py

Training and testing a version of the bot:

To train and test a specific version (e.g. 1, 2, or 3), run:

make train-version version=1

and

make test-version version=1

Versions

version_1

CALM implementation. Business logic implemented. Referred to as "ours" in the arxiv paper.

version_2

Intent-based version, constrained to the CALM code budget. Not used in the paper, as it's too weak of a baseline.

version_3

Intent-based version with larger code budget, referred to as 'baseline' in the arxiv paper. Note that version 3 is a superset of the implementation of version_2 and many files are symbolic links to the version_2 implementation.