Robust Question Answering against Distribution Shifts with Test-Time Adaptation: An Empirical Study

This repository contains the code for our paper Robust Question Answering against Distribution Shifts with Test-Time Adaptation: An Empirical Study at Findings of EMNLP2022, which includes our proposed COLDQA benchmark and the implementations of methods for test-time adaptation (i.e., Tent, PL and OIL).

Quick Links

• Overview
• COLDQA Benchmark
  • Source Distribution
  • Target Distribution
• TTA Method: Quick Start
  • Requirements
  • Download the data
  • Obtain Base Model
  • Run TTA Method
• Main Results

Overview

A deployed question answering (QA) model can easily fail when the test data has a distribution shift compared to the training data. Robustness tuning (RT) methods have been widely studied to enhance model robustness against distribution shifts before model deployment. However, can we improve a model after deployment? To answer this question, we evaluate test-time adaptation (TTA) to improve a model after deployment, and compare it with RT methods.

Specifically,

• We first introduce a new benchmark named COLDQA for robust QA, which unifies the evaluation on cross-lingual transfer, domain adaptation and corruption robustness,
• based on which, we present OIL for real-time TTA, a novel TTA method that is inspired by imitation learning and further enhanced by casual inference.

COLDQA Benchmark

We introduce COLDQA to study robust QA under distribution shifts, which is a unified evaluation benchmark against text Corruptions, cross languages and domain shifts.

To do the evaluation on COLDQA, a model should be firstly trained on the source distribution, then test the trained model on each subset from each target dataset.

Source Distribution

The training data for the source distribution is from SQuAD v1.1.

Source	|Train|	|Dev|
SQuAD	87,599	34,726

Target Distribution

Target Dataset	Distribution Shift	|Subset|	|Test|	Metric
NoiseQA-syn	Text Corruptions	3	1,190	EM / F1
NoiseQA-na	Text Corruptions	3	1,190	EM / F1
XQuAD	Cross Languages	11	1,190	EM / F1
MLQA	Cross Languages	7	4,517–11,590	EM / F1
HotpotQA	Domain Shifts	1	5,901	EM / F1
NaturalQA	Domain Shifts	1	12,836	EM / F1
NewsQA	Domain Shifts	1	4,212	EM / F1
SearchQA	Domain Shifts	1	16,980	EM / F1
TriviaQA	Domain Shifts	1	7,785	EM / F1

TTA Method: Quick Start

TTA adapts a source model with test-time data from a target distribution.

We will show how to run TTA method with our code repository in this part.

Requirements

You should run the following script to install the dependencies first.

pip install --user .
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

(To be notated, the results in our paper were obtained by training on the GPU of NVIDIA A100.)

Download the data

To this end, follow these steps:

1. create a download folder with mkdir -p download in the root of this project,
2. download the datasets from here, unzip them, and put the unzipped files under the download folder.

Obtain the Base Model

To evaluate the model generalization on COLDQA, we firstly need to train a source model with the source training data.

In the paper, we adapt the xlm-roberta-base/large, xlm-roberta-base/large with xTune as our base model. You can train the former two from here and the latter two from here.

For better reproduction, we directly provide the model trained on the source dataset for download:

• xlm-roberta-base,
• xlm-roberta-large,
• xlm-roberta-base (xTune),
• xlm-roberta-large (xTune).

Run TTA Methods

You may run by the following command:

bash scripts/run_[TASK].sh [MODEL] [DATA] [GPU] [MODEL_PATH] [SEED] [METHOD]

The predictions will be saved to the MODEL_PATH directory.

Arguments for the evaluation script are as follows:

• TASK: Distribution shift type. Choose in noiseqa, xquad, mlqa or mrqa,
• MODEL: Base model type. Choose in xlm-roberta-base or xlm-roberta-large,
• DATA: Data type. Choose in noiseQA-syn, noiseQA-na, xquad, mlqa or mrqa (should match the corresponding TASK),
• GPU: GPU ID,
• MODEL_PATH: The path of a trained base model checkpoint directory,
• SEED: Random seed,
• METHOD: TTA method. Choose in Tent, PL or OIL.

Detailed hyper-parameters can be seen in the corresponding scripts.

Main Results on COLDQA

For xlmr-base:

Method	EM (Avg)	F1 (Avg)
xlmr-base	55.11	69.21
xTune	58.54	71.94
Tent	56.24	69.68
PL	56.45	69.78
OIL	57.06	70.38
xTune+PL	58.86	71.89
xTune+OIL	59.63	72.68

For xlmr-large:

Method	EM (Avg)	F1 (Avg)
xlmr-large	58.58	73.82
xTune	61.51	76.06
Tent	54.56	70.34
PL	61.80	76.05
OIL	62.04	76.19
xTune+PL	63.73	77.01
xTune+OIL	64.57	77.93