Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance

Overview

Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance

Abstract: Models for natural language understanding (NLU) tasks often rely on the idiosyncratic biases of the dataset, which make them brittle against test cases outside the training distribution. Recently, several proposed debiasing methods are shown to be very effective in improving out-of-distribution performance. However, their improvements come at the expense of performance drop when models are evaluated on the in-distribution data, which contain examples with higher diversity. This seemingly inevitable trade-off may not tell us much about the changes in the reasoning and understanding capabilities of the resulting models on broader types of examples beyond the small subset represented in the out-of-distribution data. In this paper, we address this trade-off by introducing a novel debiasing method, called confidence regularization, which discourage models from exploiting biases while enabling them to receive enough incentive to learn from all the training examples. We evaluate our method on three NLU tasks and show that, in contrast to its predecessors, it improves the performance on out-of-distribution datasets (e.g., 7pp gain on HANS dataset) while maintaining the original in-distribution accuracy.

The repository contains the code to reproduce our work in debiasing NLU models without in-distribution degradation. We provide 2 runs of experiment that are shown in our paper:

  1. Debias MNLI model from syntactic bias and evaluate on HANS as the out-of-distribution data.
  2. Debias MNLI model from hypothesis-only bias and evaluate on MNLI-hard sets as the out-of-distribution data.

Requirements

The code requires python >= 3.6 and pytorch >= 1.1.0.

Additional required dependencies can be found in requirements.txt. Install all requirements by running:

pip install -r requirements.txt

Data

Our experiments use MNLI dataset version provided by GLUE benchmark. Download the file from here, and unzip under the directory ./dataset Additionally download the following files here for evaluating on hard/easy splits of both MNLI dev and test sets. The dataset directory should be structured as the following:

└── dataset 
    └── MNLI
        ├── train.tsv
        ├── dev_matched.tsv
        ├── dev_mismatched.tsv
        ├── dev_mismatched.tsv
        ├── dev_matched_easy.tsv
        ├── dev_matched_hard.tsv
        ├── dev_mismatched_easy.tsv
        ├── dev_mismatched_hard.tsv
        ├── multinli_hard
        │   ├── multinli_0.9_test_matched_unlabeled_hard.jsonl
        │   └── multinli_0.9_test_mismatched_unlabeled_hard.jsonl
        ├── multinli_test
        │   ├── multinli_0.9_test_matched_unlabeled.jsonl
        │   └── multinli_0.9_test_mismatched_unlabeled.jsonl
        └── original

Running the experiments

For each evaluation setting, use the --mode and --which_bias arguments to set the appropriate loss function and the type of bias to mitigate (e.g, hans, hypo).

To reproduce our result on MNLI ⮕ HANS, run the following:

cd src/
CUDA_VISIBLE_DEVICES=6 python train_distill_bert.py \
    --output_dir ../checkpoints/hans/bert_confreg_lr5_epoch3_seed444 \
    --do_train --do_eval --mode smoothed_distill \
    --seed 444 --which_bias hans

For the MNLI ⮕ hard splits, run the following:

cd src/
CUDA_VISIBLE_DEVICES=10 python train_distill_bert.py \
    --output_dir ../checkpoints/hypo/bert_confreg_lr5_epoch3_seed111 \
    --do_train --do_eval --mode smoothed_distill \
    --seed 111 --which_bias hypo

Expected results

Results on the MNLI ⮕ HANS setting:

Mode Seed MNLI-m MNLI-mm HANS avg.
None 111 84.57 84.72 62.04
conf-reg 111 84.17 85.02 69.62

Results on the MNLI ⮕ Hard-splits setting:

Mode Seed MNLI-m MNLI-mm MNLI-m hard MNLI-mm hard
None 111 84.62 84.71 76.07 76.75
conf-reg 111 85.01 84.87 78.02 78.89

Contact

Contact person: Ajie Utama, [email protected]

https://www.ukp.tu-darmstadt.de/

Please reach out to us for further questions or if you encounter any issue. You can cite this work by the following:

@InProceedings{UtamaDebias2020,
  author    = {Utama, P. Ajie and Moosavi, Nafise Sadat and Gurevych, Iryna},
  title     = {Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance},
  booktitle = {Proceedings of the 58th Conference of the Association for Computational Linguistics},
  month     = jul,
  year      = {2020},
  publisher = {Association for Computational Linguistics}
}

Acknowledgement

The code in this repository is build on the implementation of debiasing method by Clark et al. The original version can be found here

Owner
Ubiquitous Knowledge Processing Lab
Ubiquitous Knowledge Processing Lab
InsTrim: Lightweight Instrumentation for Coverage-guided Fuzzing

InsTrim The paper: InsTrim: Lightweight Instrumentation for Coverage-guided Fuzzing Build Prerequisite llvm-8.0-dev clang-8.0 cmake = 3.2 Make git cl

75 Dec 23, 2022
Codes for SIGIR'22 Paper 'On-Device Next-Item Recommendation with Self-Supervised Knowledge Distillation'

OD-Rec Codes for SIGIR'22 Paper 'On-Device Next-Item Recommendation with Self-Supervised Knowledge Distillation' Paper, saved teacher models and Andro

Xin Xia 11 Nov 22, 2022
MAterial del programa Misión TIC 2022

Mision TIC 2022 Esta iniciativa, aparece como respuesta frente a los retos de la Cuarta Revolución Industrial, y tiene como objetivo la formación de 1

6 May 25, 2022
Data, model training, and evaluation code for "PubTables-1M: Towards a universal dataset and metrics for training and evaluating table extraction models".

PubTables-1M This repository contains training and evaluation code for the paper "PubTables-1M: Towards a universal dataset and metrics for training a

Microsoft 365 Jan 04, 2023
🌊 Online machine learning in Python

In a nutshell River is a Python library for online machine learning. It is the result of a merger between creme and scikit-multiflow. River's ambition

OnlineML 4k Jan 02, 2023
torchlm is aims to build a high level pipeline for face landmarks detection, it supports training, evaluating, exporting, inference(Python/C++) and 100+ data augmentations

💎A high level pipeline for face landmarks detection, supports training, evaluating, exporting, inference and 100+ data augmentations, compatible with torchvision and albumentations, can easily instal

DefTruth 142 Dec 25, 2022
A Python framework for conversational search

Chatty Goose Multi-stage Conversational Passage Retrieval: An Approach to Fusing Term Importance Estimation and Neural Query Rewriting Installation Ma

Castorini 36 Oct 23, 2022
ConformalLayers: A non-linear sequential neural network with associative layers

ConformalLayers: A non-linear sequential neural network with associative layers ConformalLayers is a conformal embedding of sequential layers of Convo

Prograf-UFF 5 Sep 28, 2022
Consensus Learning from Heterogeneous Objectives for One-Class Collaborative Filtering

Consensus Learning from Heterogeneous Objectives for One-Class Collaborative Filtering This repository provides the source code of "Consensus Learning

SeongKu-Kang 6 Apr 29, 2022
RETRO-pytorch - Implementation of RETRO, Deepmind's Retrieval based Attention net, in Pytorch

RETRO - Pytorch (wip) Implementation of RETRO, Deepmind's Retrieval based Attent

Phil Wang 556 Jan 04, 2023
Code accompanying the paper "Knowledge Base Completion Meets Transfer Learning"

Knowledge Base Completion Meets Transfer Learning This code accompanies the paper Knowledge Base Completion Meets Transfer Learning published at EMNLP

14 Nov 27, 2022
Code for Two-stage Identifier: "Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition"

Code for Two-stage Identifier: "Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition", accepted at ACL 2021. For details of the model and experiments, please see our paper.

tricktreat 87 Dec 16, 2022
multimodal transformer

This repo holds the code to perform experiments with the multimodal autoregressive probabilistic model Transflower. Overview of the repo It is structu

Guillermo Valle 68 Dec 13, 2022
Official Implementation of Swapping Autoencoder for Deep Image Manipulation (NeurIPS 2020)

Swapping Autoencoder for Deep Image Manipulation Taesung Park, Jun-Yan Zhu, Oliver Wang, Jingwan Lu, Eli Shechtman, Alexei A. Efros, Richard Zhang UC

449 Dec 27, 2022
LIMEcraft: Handcrafted superpixel selectionand inspection for Visual eXplanations

LIMEcraft LIMEcraft: Handcrafted superpixel selectionand inspection for Visual eXplanations The LIMEcraft algorithm is an explanatory method based on

MI^2 DataLab 4 Aug 01, 2022
Code for Transformer Hawkes Process, ICML 2020.

Transformer Hawkes Process Source code for Transformer Hawkes Process (ICML 2020). Run the code Dependencies Python 3.7. Anaconda contains all the req

Simiao Zuo 111 Dec 26, 2022
Practical tutorials and labs for TensorFlow used by Nvidia, FFN, CNN, RNN, Kaggle, AE

TensorFlow Tutorial - used by Nvidia Learn TensorFlow from scratch by examples and visualizations with interactive jupyter notebooks. Learn to compete

Alexander R Johansen 1.9k Dec 19, 2022
Christmas face app for Decathlon xmas coding party!

Christmas Face Application Use this library to create the perfect picture for your christmas cards! Done by Hasib Zunair, Guillaume Brassard and Samue

Hasib Zunair 4 Dec 20, 2021
[Machine Learning Engineer Basic Guide] 부스트캠프 AI Tech - Product Serving 자료

Boostcamp-AI-Tech-Product-Serving 부스트캠프 AI Tech - Product Serving 자료 Repository 구조 part1(MLOps 개론, Model Serving, 머신러닝 프로젝트 라이프 사이클은 별도의 코드가 없으며, part

Sung Yun Byeon 269 Dec 21, 2022
Implementation of paper "Towards a Unified View of Parameter-Efficient Transfer Learning"

A Unified Framework for Parameter-Efficient Transfer Learning This is the official implementation of the paper: Towards a Unified View of Parameter-Ef

Junxian He 216 Dec 29, 2022