The Transfer Learning in Dialogue Benchmarking Toolkit

This repo contains data and code used in FETA: A Benchmark for Few-Sample Task Transfer in Open-Domain Dialogue, presented at EMNLP 2022

The repo can also be utilized for many more research scenarios, including:

• Multi-Task Learning
• In-Context Task Transfer
• Continual Learning
• Generalizability of pre-training datasets and model architectures

This repo is also the starter code for the FETA Benchmark Challenge!

The FETA Benchmark Challenge is being hosted at the 5th Workshop on NLP For Conversational AI (co-located with ACL 2023).
The mission of the FETA challenge is to encourage the development and evaluation of new approaches to task-transfer with limited in-domain data.
Specifically, FETA focuses on the dialogue domain due to interests in empowering human-machine communication through natural language.

For more details on the FETA challenge, see the FETA README.

Overview

TLiDB is a tool used to benchmark methods of transfer learning in conversational AI. TLiDB can easily handle domain adaptation, task transfer, multitasking, continual learning, and other transfer learning settings. TLiDB maintains a unified json format for all datasets and tasks, easing the new code necessary for new datasets and tasks. We highly encourage community contributions to the project.

The main features of TLiDB are:

1. Dataset class to easily load a dataset for use across models
2. Unified metrics to standardize evaluation across datasets
3. Extensible Model and Algorithm classes to support fast prototyping

Installation

Requirements

• python>=3.6
• torch>=1.10
• nltk>=3.6.5
• scikit-learn>=1.0
• transformers>=4.11.3
• sentencepiece>=0.1.96
• bert-score==0.3.11

To use TLiDB, you can simply install via pip:

pip install tlidb

OR, you can install TLiDB from source. This is recommended if you want to edit or contribute:

git clone [email protected]:alon-albalak/TLiDB.git
cd TLiDB
pip install -e .

How to use TLiDB

TLiDB can be used from the command line or as a python command. If you have installed the package from source, we highly recommend running commands from inside the tlidb/examples/ directory.

Quick Start

For a very simple set up, you can use the following commands.

• From command line:

tlidb --source_datasets Friends --source_tasks emory_emotion_recognition --target_datasets Friends --target_tasks reading_comprehension --do_train --do_finetune --do_eval --eval_best --model_config bert --few_shot_percent 0.1

• As python command (only if installed from source):

cd examples
python3 run_experiment.py --source_datasets Friends --source_tasks emory_emotion_recognition --target_datasets Friends --target_tasks reading_comprehension --do_train --do_finetune --do_eval --eval_best --model_config bert --few_shot_percent 0.1

Detailed Usage

TLiDB has 2 main folders of interest:

• tlidb/examples
• tlidb/TLiDB

tlidb/examples/ is recommended for use if you would like to utilize our training scripts. It contains sample code for models, learning algorithms, and sample training scripts. For detailed examples, see the Examples README.

tlidb/TLiDB/ holds the code related to data (datasets, dataloaders, metrics, etc.). If you are interested in utilizing our datasets and metrics but would like to train models using your own training scripts, take a look at the example usage in TLiDB README.

Folder descriptions:

• tlidb/TLiDB is the folder holding the code for data handling
  • tlidb/TLiDB/data_loaders contains code for data_loaders
  • tlidb/TLiDB/data is the destination folder for downloaded datasets (if installed from source, otherwise data is in .cache/tlidb/data)
  • tlidb/TLiDB/datasets contains code for dataset loading and preprocessing
  • tlidb/TLiDB/metrics contains code for loss and evaluation metrics
  • tlidb/TLiDB/utils contains utility files
• tlidb/examples contains sample code for training and evaluating models
  • tlidb/examples/algorithms contains code which trains and evaluates a model
  • tlidb/examples/models contains code to define a model
  • tlidb/examples/configs contains code for model configurations
• /dataset_preprocessing is for reproducability purposes. It contains scripts used to preprocess the TLiDB datasets from their original form into the standardized TLiDB format

Comments, Questions, and Feedback

If you find issues, please open an issue here.

If you have dataset or model requests, please add a new discussion here.

We encourage outside contributions to the project!

Citation

If you use the FETA datasets in your work, please cite the FETA paper:

@inproceedings{albalak-etal-2022-feta,
    title = "{FETA}: A Benchmark for Few-Sample Task Transfer in Open-Domain Dialogue",
    author = "Albalak, Alon  and
      Tuan, Yi-Lin  and
      Jandaghi, Pegah  and
      Pryor, Connor  and
      Yoffe, Luke  and
      Ramachandran, Deepak  and
      Getoor, Lise  and
      Pujara, Jay  and
      Wang, William Yang",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.751",
    pages = "10936--10953",
    abstract = "Task transfer, transferring knowledge contained in related tasks, holds the promise of reducing the quantity of labeled data required to fine-tune language models. Dialogue understanding encompasses many diverse tasks, yet task transfer has not been thoroughly studied in conversational AI. This work explores conversational task transfer by introducing FETA: a benchmark for FEw-sample TAsk transfer in open-domain dialogue.FETA contains two underlying sets of conversations upon which there are 10 and 7 tasks annotated, enabling the study of intra-dataset task transfer; task transfer without domain adaptation. We utilize three popular language models and three learning algorithms to analyze the transferability between 132 source-target task pairs and create a baseline for future work.We run experiments in the single- and multi-source settings and report valuable findings, e.g., most performance trends are model-specific, and span extraction and multiple-choice tasks benefit the most from task transfer.In addition to task transfer, FETA can be a valuable resource for future research into the efficiency and generalizability of pre-training datasets and model architectures, as well as for learning settings such as continual and multitask learning.",
}

If you use TLiDB in your work, please cite the repository:

@software{Albalak_The_Transfer_Learning_2022,
author = {Albalak, Alon},
doi = {10.5281/zenodo.6374360},
month = {3},
title = {{The Transfer Learning in Dialogue Benchmarking Toolkit}},
url = {https://github.com/alon-albalak/TLiDB},
version = {1.0.0},
year = {2022}
}

Acknowledgements

The design of TLiDB was based the wilds project, and the Open Graph Benchmark.