DAMP: Doubly Aligned Multilingual Parser for Task-Oriented Dialogue

This repository contains code to evaluate HuggingFace Seq2Seq models on the TOP family of datasets along with Constrained Adversarial Alignment.

Feel free to contact William Held with any questions at wheld3 [@] gatech edu.

For a description of methods, as well as reported results combining AMBER with pre-training alignment - see the [Paper]

Pre-Requisites

pip install -r requirements.txt

Reproducing MT5 Results with Pretrained Models

mT5 Small

Non-Adversarial: bash eval_t5.sh Adverserial: bash eval_t5_adv.sh

mT5 Base (Requires 2 GPUs)

Non-Adversarial: bash eval_t5_base.sh Non-Adversarial: bash eval_t5_base_adv.sh

Training Models From Scratch

mT5 Small

Non-Adversarial: bash train_t5.sh Adverserial: bash train_t5_adv.sh

mT5 Base (Requires 2 GPUs w/ 12GB RAM Each)

Non-Adversarial: bash train_t5_base.sh Non-Adversarial: bash train_t5_base_adv.sh

Constrained Adversarial Alignment Implementation

alignment_mixin.py