# featurelevel_debiased_natural_language_understanding__36f7f008.pdf

Feature-Level Debiased Natural Language Understanding

Yougang Lyu1, Piji Li2, Yechang Yang1, Maarten de Rijke3, Pengjie Ren1

Yukun Zhao1,4, Dawei Yin4, Zhaochun Ren1*

1School of Computer Science and Technology, Shandong University, Qingdao, China 2College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China 3University of Amsterdam, Amsterdam, The Netherlands 4Baidu Inc., Beijing, China {youganglyu, yyc002}@mail.sdu.edu.cn, pjli@nuaa.edu.cn, m.derijke@uva.nl, jay.ren@outlook.com, zhaoyukun02@baidu.com, yindawei@acm.org, zhaochun.ren@sdu.edu.cn

Natural language understanding (NLU) models often rely on dataset biases rather than intended task-relevant features to achieve high performance on specific datasets. As a result, these models perform poorly on datasets outside the training distribution. Some recent studies address this issue by reducing the weights of biased samples during the training process. However, these methods still encode biased latent features in representations and neglect the dynamic nature of bias, which hinders model prediction. We propose an NLU debiasing method, named debiasing contrastive learning (DCT), to simultaneously alleviate the above problems based on contrastive learning. We devise a debiasing, positive sampling strategy to mitigate biased latent features by selecting the least similar biased positive samples. We also propose a dynamic negative sampling strategy to capture the dynamic influence of biases by employing a bias-only model to dynamically select the most similar biased negative samples. We conduct experiments on three NLU benchmark datasets. Experimental results show that DCT outperforms state-of-the-art baselines on out-of-distribution datasets while maintaining in-distribution performance. We also verify that DCT can reduce biased latent features from the model s representations.

1 Introduction Pre-trained language models such as BERT (Devlin et al. 2019) have achieved impressive performance on many NLU benchmarks, such as natural language inference (NLI) (Bowman et al. 2015; Williams, Nangia, and Bowman 2018) and fact verification (Thorne et al. 2018). However, recent studies have shown that these models tend to leverage dataset biases instead of intended task-relevant features (Mc Coy, Pavlick, and Linzen 2019; Schuster et al. 2019; Du et al. 2022). For example, Gururangan et al. (2018) find that NLU models rely on the spurious association between negative words (e.g., nobody, no, never and nothing) and contradiction labels for prediction in NLI datasets, leading to low accuracy on outof-distribution datasets that lack spurious associations. To mitigate bias in training datasets, recent NLU debiasing methods attempt to train more robust models. Three prevailing debiasing methods exist in NLU: (i) example

*Corresponding author. Copyright 2023, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.

(a) Different debiasing methods

(b) Different training epochs

Figure 1: Probing accuracy for three types of biased features (Overlap, Subsequence and Negwords) with different methods and different training epochs on the MNLI dataset. (a) Recent NLU debiasing methods (Reweight, POE, and Conf-reg) have a higher probing accuracy of biased features compared to BERT-base. (b) Biased features probing accuracy of BERT-base changes dynamically during the training process.

reweighting (Reweight) (Schuster et al. 2019), (ii) product-of experts (POE) (Clark, Yatskar, and Zettlemoyer 2019; He, Zha, and Wang 2019; Mahabadi, Belinkov, and Henderson 2020), and (iii) confidence regularization (Conf-reg) (Utama, Moosavi, and Gurevych 2020a). These debiasing methods encourage the model to pay less attention to the biased examples, which forces it to learn harder samples to improve out-of-distribution (OOD) performance. From the perspective of debiasing NLU, two main challenges remain. Both concern biased features, that is, features that have spurious correlations with the label, e.g., negative words in input sentences in NLU tasks. First, existing NLU debiasing methods still encode biased latent features in representations. We follow Mendelson and Belinkov (2021) to use probing tasks for several types of bias (Overlap, Subsequence and Negwords) to verify whether biased latent features have been removed from representations. The probing task for bias is to predict whether a sample is biased based on the model s representation in terms of probing accuracy, the accuracy of the probing task. Higher probing accuracy of bias means that the model s representation contains more biased features. Fig. 1(a) shows that existing debiasing methods have a higher probing accuracy of three types of biased latent features from

The Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI-23)

the representations than the fine-tuned BERT-base. These results illustrate that existing debiasing methods do not improve OOD performance by reducing biased features and modeling intended task-relevant features, but by adjusting the conditional probability of labels given biased features. Since these debiasing methods only adjust the conditional probability of labels given biased features, they improve OOD performance at the cost of degrading in-distribution (ID) performance. This poses a feature-level debiasing challenge to debiasing approaches. Second, existing NLU debiasing methods neglect the dynamic influence of bias. The number of biased latent features in a representation changes during the training process. Since the model predicts the label based on the representation, biased features dynamically influence model prediction during the training process. In Fig. 1(b) we examine three types of bias and find that the probing accuracy of biases changes during training. Different types of biased features have a different influence on model prediction during training. It is important to reduce biased features that have the greatest influence on model prediction at different training epochs. This poses a challenge in capturing the dynamic influence of bias. To tackle the above challenges, we propose a novel debiasing method, debiasing contrastive learning (DCT). The main idea of DCT is to encourage positive examples with least similar bias to be closer and negative examples with most similar bias to be apart at the feature-level. DCT consists of two strategies: (i) a debiasing, positive sampling strategy to mitigate biased latent features, and (ii) a dynamic negative sampling strategy to capture the dynamic influence of biased features. As to the first strategy, DCT s debiasing, positive sampling strategy selects the least similar biased positive samples from the debiasing dataset. We filter debiasing samples from the training set where the bias-only model has high confidence but incorrect predictions. As the bias-only model relies only on biased features to make predictions, the debiasing dataset contains biased features, but the correlation between biased features and labels differs from the dominant spurious correlation in the training dataset. As to the second strategy, DCT s dynamic negative sampling strategy uses the bias-only model to dynamically select the most similar biased negative sample during the training process. Furthermore, we adopt momentum contrast (He et al. 2020) to establish a massive queue for saving representations dynamically. We conduct experiments on three NLU benchmark datasets to evaluate bias extractability and debiasing performance of our proposed method. DCT outperforms stateof-the-art baselines on OOD datasets and maintains ID performance by reducing multiple types of biased features in a model s representations. To sum up, our contributions are as follows: To the best of our knowledge, we are the first to focus on feature-level debiasing and modeling the dynamic influence of bias in NLU tasks at the same time. We propose a novel debiasing method, named DCT, which uses contrastive learning and combines a debiasing, positive sampling strategy and a dynamic negative sampling strategy to reduce biased latent features and capture the

dynamic influence of biases. Experiments on three NLU benchmark datasets show that DCT reduces biased latent features in the model s representation and outperforms state-of-the-art baselines on OOD datasets while maintaining ID performances.1

2 Related Work 2.1 Dataset Bias Exploiting dataset biases seems easier for deep neural networks than learning the intended task-relevant features. (Geirhos et al. 2020; Bender and Koller 2020). For instance, models can perform better than most baselines by only partially using input in NLI without capturing the semantic relationships between premises and hypothesis sentences (Gururangan et al. 2018). Similar phenomena have been observed in other tasks, e.g., visual question-answering (Agrawal, Batra, and Parikh 2016), reading comprehension (Kaushik and Lipton 2018), and paraphrase identification (Zhang, Baldridge, and He 2019). Prior work has constructed challenge datasets consisting of counter examples to superficial cues that deep neural networks might adopt (Jia and Liang 2017; Glockner, Shwartz, and Goldberg 2018; Naik et al. 2018; Mc Coy, Pavlick, and Linzen 2019). When models are evaluated on these challenge datasets, their performance often drops to the same as the random baseline (Gururangan et al. 2018; Schuster et al. 2019). Therefore, there is a clear need for methods that are tailored to address NLU dataset biases.

2.2 Debiasing NLU Methods Several studies aim to mitigate dataset bias by improving dataset construction techniques. For example, Zellers et al. (2019); Sakaguchi et al. (2021) reduce biased patterns in datasets with adversarial filtering; Nie et al. (2020); Kaushik, Hovy, and Lipton (2020) adopt a dynamic, human-in-the-loop data collection technique; Min et al. (2020); Schuster, Fisch, and Barzilay (2021) use adversarial samples to augment the training dataset; and Wu et al. (2022); Ross et al. (2022) train data generators to generate debiased datasets. A complementary line of work trains more robust models with alternative learning algorithms, such as product-of-experts (POE) (Clark, Yatskar, and Zettlemoyer 2019; He, Zha, and Wang 2019; Mahabadi, Belinkov, and Henderson 2020), confidence regularization (Conf-reg) (Utama, Moosavi, and Gurevych 2020a; Du et al. 2021), and example reweighting (Reweight) (Schuster et al. 2019). These algorithms can be formalized as two-stage frameworks; the first step is to train a bias-only model, either automatically (Utama, Moosavi, and Gurevych 2020b; Sanh et al. 2021; Ghaddar et al. 2021) or using prior knowledge about the bias (Clark, Yatskar, and Zettlemoyer 2019; He, Zha, and Wang 2019; Belinkov et al. 2019b,a); at the second stage, the output of the bias-only model is used to adjust the loss function of the debiased model. However, these debiasing methods make biased features more extractable from the model representations (Mendelson and Belinkov 2021; Du et al. 2022). Instead, we aim to use contrastive learning to dynamically push the NLU

1The code is available at https://github.com/youganglyu/DCT

Task-relevant feature

Push Training Most similar biased negative Least similar biased positive

Biased feature

Task-relevant feature

Biased feature Biased feature

Task-relevant feature

Biased feature

(a) Cross entropy (b) Cross entropy + Debiasing contrastive learning (Ours)

Task-relevant feature

Training sample with class A

Training sample with class B

OOD test sample with class A

OOD test sample with class B

Classification plane

Figure 2: (a) Models fine-tuned with cross-entropy use biased features to predict. (b) Models fine-tuned with cross-entropy and DCT reduce biased features.

model to reduce biased features and capture intended taskrelevant features. This enables the NLU model to improve OOD performance while maintaining ID performance.

2.3 Contrastive Learning The main idea of contrastive learning is to encourage the representation of similar samples to be close and different samples to be apart (Hadsell, Chopra, and Le Cun 2006; Chen et al. 2020). Contrastive learning has been used to improve improve in-distribution performance in computer vision (CV; He et al. 2020; Chen et al. 2020) and natural language processing (NLP; Giorgi et al. 2021; Wang et al. 2021). In a self-supervised framework, positive (i.e., similar) samples can be generated by data augmentation of the anchor sample, and negative (i.e., different) samples can be obtained from the same batch (Gao, Yao, and Chen 2021) or from a memory bank/queue that saves the representation of previous samples (Wu et al. 2018; Chen et al. 2020). In a supervised framework, positive samples belong to the same class while negative samples belong to a different class (Khosla et al. 2020; Gunel et al. 2021; Li et al. 2021). These methods focus on improving ID performance, while we aim to use contrastive learning to reduce biased latent features and improve OOD performance.

In this section, we detail the DCT method. First, we formulate our research problem. Then, we introduce the overall framework of debiasing contrastive learning. Next, we introduce the debiasing, positive sampling strategy and describe the dynamic negative sampling strategy. Finally, a training process with momentum contrast for DCT is explained.

3.1 Problem Formulation

Following (Clark, Yatskar, and Zettlemoyer 2019; He, Zha, and Wang 2019; Mahabadi, Belinkov, and Henderson 2020; Utama, Moosavi, and Gurevych 2020a), we formulate NLU tasks as a general classification problem. We denote a training dataset as D consisting of N examples {xi, yi}N i=1, where xi X is the input data, yi Y is the target label, |Y| = K

is the number of the classes. For each input instance x, we assume that the features of x can be divided into intended task-relevant features xt and biased features xb, where xt have invariant relations with the label y and xb have spurious relations with the label y. The random variables of x, y, xb and xt are respectively denoted as X, Y , Xb and Xt. Our goal is to train a debiasing model fd to capture PD(Y |Xt) by reducing biased features Xb, which performs better on OOD datasets and maintains ID performance.

3.2 Debiasing Contrastive Learning DCT aims to pull the least similar biased positive samples closer to each other and push the most similar biased negative samples apart at the feature-level, as illustrated in Fig. 2(b). To accomplish this, we rewrite the contrastive loss to obtain the debiasing contrastive learning loss as follows:

1 |Sp i | P

log exp(Φd(xi) Φd (xj)/τ) P

xk Sn i /{xj } exp(Φd(xi) Φd (xk)/τ)

where Φd( ) refers to the debias encoder, Φd( ) denotes to the momentum encoder, |Sp i | is the size of the debiasing, positive sample set Sp i for xi, Sn i denotes the negative sample set for xi, and τ is a scalar temperature parameter. In the following subsections, we detail the debiasing, positive sampling strategy and the dynamic negative sampling strategy.

Task-relevant feature

Biased feature

Dominant biased training sample

Debiasing training sample

Class A Class B

Classification plane Figure 3: The bias-only model is used to filter the debiasing dataset from the original training dataset.

3.3 Debiasing Positive Sampling To mitigate biased latent features, we employ a bias-only model to filter the debias dataset from the training dataset and sample least similar biased positive samples from the debias dataset, i.e., we train a bias-only model fb to approximate PD(Y |Xb) which makes predictions only based on biased features. Following Sanh et al. (2021), we use a limited capacity weak learner as the bias-only model, which is trained on the full training dataset. As sketched in Fig. 3, if the bias-only model is particularly confident in predicting a sample but predicts incorrectly, it is likely that the sample contains biased features, but the correlation between biased features and labels differs from the dominant spurious correlation in the training dataset. Given a training example {xi, yi}, we assume the output of bias-only model fb to be bi = bi,1, bi,2, . . . , bi,K . Based on the probabilistic distributions bi, we filter the debiasing dataset from the training dataset, so we have:

Ddebias = {{xi, yi}|bi,c λ yi,c = 0}, (2)

where c is the predicted class by the bias-only model, bi,c denotes the scalar probability value of the class c, yi,c refers to the ground truth of class c for xi, and λ is a scalar threshold. To construct the positive sample set Sp i for xi in Eq. 1, we select positives that are least similar to xi in the debiasing dataset by L2 distance. Additionally, we incorporate the debiasing dataset into the training dataset to generate a sufficient number of positive pairs.

3.4 Dynamic Negative Sampling To capture the dynamic influence of bias, we employ the bias-only model to dynamically select the most similar biased negative sample that is closest to the anchor sample xi. Based on the checkpoints of the bias-only model, we apply the encoder of the bias-only model of each epoch to all the samples and dynamically retrieve the most similar biased negative sample set, that is:

Sdn i,k = {xj|yj = yi arg minj(d(Φk b(xi), Φk b(xj)))}, (3)

where Sdn i,k denotes the set of most similar biased negative samples of xi for the k-th epoch, Φk b( ) is the bias-only model encoder for the k-th epoch, and d( ) refers to the L2 distance function. To construct the negative sample set Sn i for xi in Eq. 1, we combine Sdn i,k and other negative samples in the momentum contrast queue.

3.5 Training with Momentum Contrast To leverage a large number of positive and negative samples, we adopt momentum contrast (He et al. 2020) to build a massive queue for dynamically saving representations. In order to maintain the representation consistency in the queue, the momentum contrast framework requires two encoders, a debias encoder and a momentum encoder. During training with the momentum contrast framework, the parameter θd in the debias encoder is updated by training samples and then the θd in the momentum encoder is updated by:

θd mθd + (1 m) θd , (4)

where m [0, 1) is a momentum coefficient, which keeps the consistency of sample representations in the queue. The sample representations in the queue are gradually replaced. Specifically, the sample representations encoded by the momentum encoder are added to the queue, and the oldest sample representations are removed. In each training iteration, only the parameter θd in the debias encoder is updated by back-propagation. To directly use the label information, we adopt cross entropy as part of the overall loss for training the main model:

LCE = yi log fd (xi) . (5)

The overall loss function is formally computed as:

L = (1 α) LCE + αLDCT , (6)

where α is a scalar weighting hyperparameter.

4 Experiments 4.1 Research Questions We conduct experiments on different NLU tasks to answer the following research questions: (RQ1) Does the proposed DCT method reduce multiple types of biased features simultaneously? (RQ2) How does the proposed DCT perform on ID and OOD datasets compared to state-of-the-art baselines? (RQ3) How do strategies and hyperparameters affect ID and OOD performances of DCT?

4.2 Datasets We use three NLU benchmark datasets in our experiments:

MNLI The MNLI dataset (Williams, Nangia, and Bowman 2018) contains pairs of premise and hypothesis sentences labeled as entailment, neutral, and contradiction. We test models trained on MNLI against the challenge dataset HANS (Mc Coy, Pavlick, and Linzen 2019). It contains examples of high overlap between premises and hypothetical sentences but are labeled as contradiction. Since the overlapping feature is correlated with label entailment in MNLI, models trained directly on MNLI tend to perform poorly on HANS. SNLI The SNLI dataset (Bowman et al. 2015) contains pairs of premise and hypothesis sentences labeled as entailment, neutral, and contradiction. Following Utama et al. (2021), we evaluate models trained on SNLI against the long and short subsets of the Scramble Test challenge set (Dasgupta et al. 2018). It contains samples that are changed the word order against the overlap bias in SNLI dataset. FEVER The FEVER dataset (Thorne et al. 2018) contains pairs of claim and evidence sentences labeled as either support, not-enough-information, or refute. We follow Schuster et al. (2019) to process and split the dataset.2 FEVER models rely on the claim-only bias, where specific words in the claim are often associated with target label. We evaluate models trained on FEVER against Fever Symmetric datasets (Schuster et al. 2019) (version 1 and 2), which were manually constructed to reduce claim-only bias.

2https://github.com/Tal Schuster/Fever Symmetric

Overlap Subsequence Neg Words

Method Compression Acc. Compression Acc. Compression Acc.

BERT-base 3.29 0.16 88.84 1.22 3.21 0.24 90.91 2.53 2.44 0.12 85.35 0.93

Reweight 3.68 0.12 90.39 0.79 3.66 0.11 92.21 2.30 2.53 0.11 87.06 0.22 POE 3.68 0.10 92.41 1.27 3.82 0.17 94.10 1.97 2.47 0.09 86.70 0.85 Conf-reg 4.34 0.20 93.28 0.69 4.03 0.18 94.54 1.82 2.70 0.11 89.24 0.27 DCT 2.79 0.15 85.35 0.84 2.76 0.13 88.96 1.77 1.88 0.12 80.31 0.35

Table 1: Results of probing for Overlap, Subsequence, and Neg Words on MNLI. Acc is the probing accuracy of biases. Note that lower compression scores and probing accuracy represent lower extractability of biased features in the model representation.

Overlap Subsequence Neg Words

Method Compression Acc. Compression Acc. Compression Acc.

BERT-base 4.39 0.17 93.82 0.99 5.10 0.23 94.96 1.44 4.08 0.36 93.26 0.66

Reweight 4.72 0.28 92.79 1.05 5.44 0.50 95.22 1.62 4.17 0.29 93.33 0.53 POE 4.78 0.14 93.22 0.47 5.14 0.15 94.34 1.18 4.28 0.22 94.12 0.52 Conf-reg 5.20 0.19 94.81 0.46 5.67 0.16 93.45 2.86 4.78 0.29 95.70 0.49 DCT 3.14 0.10 90.63 0.51 3.44 0.15 91.86 2.19 2.25 0.14 85.17 0.75

Table 2: Results of probing for Overlap, Subsequence and Neg Words on SNLI. Notational conventions are the same as in Table 1.

Method Compression Acc.

BERT-base 2.57 0.08 81.82 0.75

Reweight 2.68 0.12 83.72 1.88 POE 2.77 0.05 84.60 1.08 Conf-reg 2.82 0.06 82.47 0.90 DCT 2.21 0.08 78.27 1.04

Table 3: Results of probing for Neg Words on FEVER. The notation here is consistent with Table 1.

4.3 Baselines and Evaluation Metrics We compare DCT with three state-of-the-art debiasing methods: (i) Example reweighting (Reweight) (Schuster et al. 2019) adjusts the importance of each training instance by computing the importance weight. The weight scalar for each training instance xi is computed as 1 bi,g, where bi,g is the probability of the bias-only model predicting the gold label. (ii) Product-of-experts (POE) (Clark, Yatskar, and Zettlemoyer 2019; He, Zha, and Wang 2019; Mahabadi, Belinkov, and Henderson 2020) trains a debiased model by ensembling with the bias-only model. (iii) Confidence regularization (Conf-reg) (Utama, Moosavi, and Gurevych 2020a) regularizes model confidence on biased training examples. Conf-reg uses a self-distillation training objective and scales the teacher model output by the bias-only model s output. To measure the extractability of biased features in the model s representation, we follow Mendelson and Belinkov (2021) to use compression scores and probing accuracy.3

The compression score is defined as compression = Lunif Lonline , where Lunif = |D| log K denotes the uniform distribution over the K labels and Lonline is the online coding proposed by Voita and Titov (2020). The probing accuracy is the accuracy (Acc.) of the probing task.

3https://github.com/technion-cs-nlp/bias-probing

To evaluate the ID and OOD performance of models, we follow existing work (Schuster et al. 2019; He, Zha, and Wang 2019; Utama et al. 2021) and employ accuracy (Acc.) on the ID and corresponding OOD datasets.

4.4 Implementation Details For the MNLI, SNLI and FEVER datasets, we train all models for 5 epochs; all models converge. The base model uses BERT-base (Devlin et al. 2019) and combines with crossentropy to fine-tune on three datasets. For debiased models, the first step is to train a bias-only model, where we follow Sanh et al. (2021) using Tiny BERT (Micheli, d Hoffschmidt, and Fleuret 2020) for modeling unknown bias. Based on the same bias-only model, we train all the above debiased models. In the training process, we adopt the Adam W (Loshchilov and Hutter 2019) optimizer as the optimizer with initial learning rate 3 10 5. Meanwhile, the temperature parameter τ, threshold λ, momentum coefficient m and scalar weighting hyperparameter α are set to 0.04, 0.6, 0.999, and 0.1. The sizes of the least similar positive samples Sp and the most similar negative samples Sdn are set to 150 and 1.

5 Experimental Results and Analysis To answer our research questions we conduct bias extractability experiments, ID and OOD experiments, and ablation studies are conducted. To directly explore the effectiveness of DCT in reducing biased latent features, we conducted visualization experiments.

5.1 Bias Extractability For RQ1, we analyze three types of bias and three datasets.

MNLI Table 1 shows results for the Overlap, Subsequence and Negwords probing tasks on MNLI. Compared to fine-tuned baseline (BERT-base), all debiasing methods except DCT increase the extractability of multiple types

MNLI (Acc.) SNLI (Acc.) FEVER (Acc.)

Method dev HANS dev Scramble dev Symm. v1 Symm. v2

BERT-base 84.16 0.23 61.22 1.17 90.61 0.15 72.74 6.87 87.06 0.57 56.53 0.78 63.84 0.83

Reweight 82.56 0.31 66.18 1.04 86.44 0.24 80.30 6.99 83.45 0.36 61.56 1.19 67.33 1.04 POE 81.62 0.18 67.27 1.21 83.69 0.33 79.51 6.42 82.23 0.52 62.19 1.65 67.36 1.54 Conf-reg 84.15 0.21 64.89 1.08 90.56 0.11 83.21 4.26 85.31 0.29 59.69 1.35 64.75 1.28 DCT 84.19 0.17 68.30 0.85 90.64 0.33 86.40 4.64 87.12 0.34 63.27 1.62 68.45 1.09

Table 4: Classification accuracy on MNLI, SNLI and FEVER.

of biases, which is demonstrated by higher compression values and higher probing accuracy of biases. Compared to the baselines, DCT has the lowest compression value and probing accuracy for multiple biases, indicating that our method DCT reduces the extractability of multiple types of biased features simultaneously on MNLI.

SNLI Table 2 shows results for the Overlap, Subsequence and Negwords probing tasks on SNLI. Compared to the fine-tuned baseline (BERT-base), all debiasing methods except DCT increase the extractability of multiple biases. Compared to the baselines, DCT has the lowest compression value and bias probing accuracy for multiple biases, indicating that it reduces the extractability of multiple types of biased features simultaneously on SNLI.

FEVER Table 3 shows results for the Negwords probing task on FEVER. Compared to fine-tuned baseline (BERTbase), all debiasing methods except DCT increase the extractability of Negwords bias. Compared to the baselines, DCT has the lowest compression value and bias probing accuracy for Negwords bias, indicating that our method DCT reduces the extractability of Negwords bias on FEVER.

5.2 ID and OOD Performance

Next, we turn to RQ2 and evaluate the in-distribution and outof-distribution performance of models on the development set and the corresponding challenge set of each dataset.

In-Distribution Performance. We evaluate the performance of models on the development sets of MNLI, SNLI and FEVER as the in-distribution (ID) performance. From the ID performance of models, we have the following observations: (i) Compared to debiasing baselines, our method DCT performs best on the ID development sets of MNLI, SNLI and FEVER. (ii) Compared to the fine-tuned BERT-base, Reweight and POE have substantially lower ID performance, Conf-reg can maintain ID performance on MNLI and SNLI, while DCT can maintain ID performance on MNLI, SNLI, and FEVER. (iii) Although Conf-reg also maintains the ID performance, it has the highest extractability of biased features. In contrast to Conf-reg, DCT has the lowest extractability of biased feature and maintains ID performance.

Out-of-Distribution Performance. We evaluate the performance of models on corresponding challenge sets of MNLI, SNLI and FEVER as the out-of-distribution (OOD) performance. We observe that: (i) Compared to the baselines, DCT performs well, both on the ID dataset and OOD dataset. For instance, on the MNLI, POE improves the average HANS ac-

MNLI (Acc.)

Method dev HANS avg.

DCT 84.19 68.30 76.25 -debiasing positive sampling 83.94 65.88 74.91 -dynamic negative sampling 83.54 65.79 74.67 -all 84.57 63.24 73.91

Table 5: Ablation studies with different strategies on MNLI.

curacy from 61.22 to 67.27 but sacrifices 2.54 points of MNLI in-distribution accuracy; Conf-reg maintains in-distribution accuracy but only improves 3.67 points on HANS. (ii) Compared to the fine-tuned BERT-base, all debiased baselines improve the OOD performance, and Conf-reg even achieves a trade-off between ID and OOD performance, but all debiased baselines improve the extractability of biased features. In contrast, DCT reduces the extractability of biased features while improving ID and OOD performances, indicating that our method DCT reduces biased latent features and learns intended task-relevant features.

5.3 Ablation Studies For RQ3, we perform ablation experiments with respect to strategies, threshold λ, the number of debiasing positive samples, and the number of dynamic negative samples.

Impact of Different Strategies. Table 5 lists our ablation experiments on MNLI and HANS to explore the effectiveness of strategies. (i) -debiasing positive sampling: we built the DCT model without debiasing dataset and positive samples are randomly sampled from the original training dataset. (ii) -dynamic negative sampling: we built DCT without the bias-only model and negative samples are randomly sampled from the original training dataset. (iii) -all: we remove the debiasing positive sampling strategy and dynamic negative sampling strategy simultaneously. DCT is degraded to the original supervised contrastive learning (Gunel et al. 2021). The results in Table 5 show that both strategies (debiasing positive sampling strategy and dynamic negative sampling strategy) enhance the OOD performances of DCT. The ID performance will be improved after removing all strategies compared to DCT. The reason is that supervised contrastive learning aims to improve the ID performance. However, the original supervised contrastive learning is not designed for solving dataset bias, so it has the lowest OOD performance.

Impact of Threshold λ. The threshold λ is defined in Eq. 2 to filter the debiased samples that are incorrectly predicted

MNLI (Acc.)

Threshold dev HANS avg.

0.4 84.12 66.27 75.20 0.5 84.15 67.13 75.64 0.6 84.19 68.30 76.25 0.7 84.17 66.81 75.49 0.8 84.25 64.91 74.58

Table 6: Parameter analysis for threshold λ.

by the bias-only model but with a high confidence above the threshold. We conduct an ablation study on the threshold by changing it from 0.4 to 0.8. From Table 6, when λ is set to 0.6, we can get the best OOD performance. When the value of λ is too small, some of the filtered debias samples are not misclassified by containing biased features, thus introducing noise into the debiasing process. Conversely, when λ is too large, the number and diversity of filtered debias samples are insufficient, thus affecting the debiasing process.

MNLI (Acc.)

|Sp| dev HANS avg.

75 83.96 66.94 75.45 100 84.22 67.15 75.69 150 84.19 68.30 76.25 500 84.18 66.14 75.16 1,000 84.12 66.82 75.47

Table 7: Parameter analysis for the number of debiasing positive samples.

Impact of the Number of Debiasing Positive Samples. Several experiments are conducted to explore the impact of the number of debiasing positive samples. The results are shown in Table 7. More positive samples for DCT do indeed lead to better OOD performance. In addition, with the growth of the number of positive samples, OOD performance is slightly degraded which is probably due to some positive samples contain similar biased features to the anchor sample.

MNLI (Acc.)

|Sdn| dev HANS avg.

0 83.54 65.79 74.67 1 84.19 68.30 76.25 5 83.66 68.25 75.96 10 84.07 68.06 76.06 20 84.11 68.15 76.13

Table 8: Parameter analysis for the number of dynamic negative samples.

Impact of the Number of Dynamic Negative Samples. To explore the impact of the number of dynamic negative samples, sufficient experiments are conducted. As shown in Table 8, the most similar biased negative sample improve the ID and OOD performances considerably, while adding more similar biased negative samples has little influence on the ID

Entailment Non-entailment

(a) Encoder of BERT-base.

Entailment Non-entailment

(b) Encoder of DCT.

Figure 4: t-SNE plots of the learned [CLS] embeddings on 200 samples from the HANS dataset, comparing BERT-base fine-tuned with cross-entropy only (a) and with our proposed DCT (b) for the NLI task. Red: samples contain overlap bias and are labeled as entailment; Yellow: samples contain overlap bias and are labeled as non-entailment.

and OOD performance. We can observe that the best ID and OOD performances can be achieved when we set |Sdn| to 1.

5.4 Visualizations

As mentioned before, the key idea of our proposed method DCT is to encourage positive examples with least similar bias to be closer and negative examples with most similar bias to be apart at the feature-level. To describe this insight intuitively, we use t-SNE to plot the [CLS] representations of the original BERT finetuned model and our debiased model using 200 data points sampled from the HANS dataset. The 100 data points sampled from HANS containing overlap bias are labeled as entailment and the other 100 data points containing overlap bias are labeled as non-entailment. As shown in Fig. 4(a), the encoder trained with only crossentropy is biased at the feature-level and thus cannot distinguish between samples with the same bias but different classes. In contrast, in Fig. 4(b), the encoder trained with DCT pushes away samples with the same overlap bias at the feature-level, so that samples with the same bias but different classes are more easier to distinguish.

6 Conclusions

We have focused on reducing biased latent features in an NLU model s representation and on capturing the dynamic influence of biased features. To tackle these challenges, we have proposed an NLU debiasing method, namely DCT. To mitigate biased latent features, we have proposed a debiasing, positive sampling strategy. To capture the dynamic influence of biased features, we have devised a dynamic negative sampling strategy to use the bias-only model to dynamically select the most similar biased negative sample during the training process. Experiments have shown that DCT improves the OOD performance while maintaining ID performance. In addition, our method reduces the extractability of multiple types of bias from an NLU model s representations. A limitation of DCT is that it is implemented only for the classification task. Our future work is to extend the proposed method to other NLU tasks that are impacted by dataset bias, e.g., named entity recognition and question answering.

Acknowledgments This work was supported by the National Key R&D Program of China with grant No. 2020YFB1406704, the Natural Science Foundation of China (62272274, 62202271, 61902219, 61972234, 62072279, 62102234, 62106105), the Natural Science Foundation of Shandong Province (ZR2021QF129), the Key Scientific and Technological Innovation Program of Shandong Province (2019JZZY010129), and the Hybrid Intelligence Center, a 10-year program funded by the Dutch Ministry of Education, Culture and Science through the Netherlands Organisation for Scientific Research, https: //hybrid-intelligence-centre.nl. All content represents the opinion of the authors, which is not necessarily shared or endorsed by their respective employers and/or sponsors.

References Agrawal, A.; Batra, D.; and Parikh, D. 2016. Analyzing the Behavior of Visual Question Answering Models. In Proceedings of EMNLP, 1955 1960. Belinkov, Y.; Poliak, A.; Shieber, S. M.; Durme, B. V.; and Rush, A. M. 2019a. Don t Take the Premise for Granted: Mitigating Artifacts in Natural Language Inference. In Proceedings of ACL, 877 891. Belinkov, Y.; Poliak, A.; Shieber, S. M.; Durme, B. V.; and Rush, A. M. 2019b. On Adversarial Removal of Hypothesisonly Bias in Natural Language Inference. In Proceedings of SEM, 256 262. Bender, E. M.; and Koller, A. 2020. Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data. In Proceedings of ACL, 5185 5198. Bowman, S. R.; Angeli, G.; Potts, C.; and Manning, C. D. 2015. A Large Annotated Corpus for Learning Natural Language Inference. In Proceedings of EMNLP, 632 642. Chen, T.; Kornblith, S.; Norouzi, M.; and Hinton, G. E. 2020. A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of ICML, 1597 1607. Clark, C.; Yatskar, M.; and Zettlemoyer, L. 2019. Don t Take the Easy Way Out: Ensemble Based Methods for Avoiding Known Dataset Biases. In Proceedings of EMNLP, 4067 4080. Dasgupta, I.; Guo, D.; Stuhlm uller, A.; Gershman, S.; and Goodman, N. D. 2018. Evaluating Compositionality in Sentence Embeddings. In Proceedings of Cog Sci. Devlin, J.; Chang, M.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of NAACL, 4171 4186. Du, M.; He, F.; Zou, N.; Tao, D.; and Hu, X. 2022. Shortcut Learning of Large Language Models in Natural Language Understanding: A Survey. Co RR, abs/2208.11857. Du, M.; Manjunatha, V.; Jain, R.; Deshpande, R.; Dernoncourt, F.; Gu, J.; Sun, T.; and Hu, X. 2021. Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU models. In Proceedings of NAACL, 915 929. Gao, T.; Yao, X.; and Chen, D. 2021. Sim CSE: Simple Contrastive Learning of Sentence Embeddings. In Proceedings of EMNLP, 6894 6910.

Geirhos, R.; Jacobsen, J.; Michaelis, C.; Zemel, R. S.; Brendel, W.; Bethge, M.; and Wichmann, F. A. 2020. Shortcut Learning in Deep Neural Networks. Nat. Mach. Intell., 2(11): 665 673. Ghaddar, A.; Langlais, P.; Rezagholizadeh, M.; and Rashid, A. 2021. End-to-End Self-Debiasing Framework for Robust NLU Training. In Proceedings of ACL, 1923 1929. Giorgi, J. M.; Nitski, O.; Wang, B.; and Bader, G. D. 2021. De CLUTR: Deep Contrastive Learning for Unsupervised Textual Representations. In Proceedings of ACL, 879 895. Glockner, M.; Shwartz, V.; and Goldberg, Y. 2018. Breaking NLI Systems with Sentences that Require Simple Lexical Inferences. In Proceedings of ACL, 650 655. Gunel, B.; Du, J.; Conneau, A.; and Stoyanov, V. 2021. Supervised Contrastive Learning for Pre-trained Language Model Fine-tuning. In Proceedings of ICLR. Gururangan, S.; Swayamdipta, S.; Levy, O.; Schwartz, R.; Bowman, S. R.; and Smith, N. A. 2018. Annotation Artifacts in Natural Language Inference Data. In Proceedings of NAACL, 107 112. Hadsell, R.; Chopra, S.; and Le Cun, Y. 2006. Dimensionality Reduction by Learning an Invariant Mapping. In Proceedings of CVPR, 1735 1742. He, H.; Zha, S.; and Wang, H. 2019. Unlearn Dataset Bias in Natural Language Inference by Fitting the Residual. In Proceedings of EMNLP, 132 142. He, K.; Fan, H.; Wu, Y.; Xie, S.; and Girshick, R. B. 2020. Momentum Contrast for Unsupervised Visual Representation Learning. In Proceedings of CVPR, 9726 9735. Jia, R.; and Liang, P. 2017. Adversarial Examples for Evaluating Reading Comprehension Systems. In Proceedings of EMNLP, 2021 2031. Kaushik, D.; Hovy, E. H.; and Lipton, Z. C. 2020. Learning the Difference that Makes a Difference With Counterfactually-Augmented Data. In Proceedings of ICLR. Kaushik, D.; and Lipton, Z. C. 2018. How Much Reading Does Reading Comprehension Require? A Critical Investigation of Popular Benchmarks. In Proceedings of EMNLP, 5010 5015. Khosla, P.; Teterwak, P.; Wang, C.; Sarna, A.; Tian, Y.; Isola, P.; Maschinot, A.; Liu, C.; and Krishnan, D. 2020. Supervised Contrastive Learning. In Proceedings of Neur IPS. Li, L.; Song, D.; Ma, R.; Qiu, X.; and Huang, X. 2021. KNNBERT: Fine-Tuning Pre-Trained Models with KNN Classifier. ar Xiv preprint ar Xiv:2110.02523 (2021). Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In Proceedings of ICLR. Mahabadi, R. K.; Belinkov, Y.; and Henderson, J. 2020. Endto-End Bias Mitigation by Modelling Biases in Corpora. In Proceedings of ACL, 8706 8716. Mc Coy, T.; Pavlick, E.; and Linzen, T. 2019. Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference. In Proceedings of ACL, 3428 3448. Mendelson, M.; and Belinkov, Y. 2021. Debiasing Methods in Natural Language Understanding Make Bias More Accessible. In Proceedings of EMNLP, 1545 1557.

Micheli, V.; d Hoffschmidt, M.; and Fleuret, F. 2020. On the Importance of Pre-training Data Volume for Compact Language Models. In Proceedings of EMNLP, 7853 7858. Min, J.; Mc Coy, R. T.; Das, D.; Pitler, E.; and Linzen, T. 2020. Syntactic Data Augmentation Increases Robustness to Inference Heuristics. In Proceedings of ACL, 2339 2352. Naik, A.; Ravichander, A.; Sadeh, N. M.; Ros e, C. P.; and Neubig, G. 2018. Stress Test Evaluation for Natural Language Inference. In Proceedings of COLING, 2340 2353. Nie, Y.; Williams, A.; Dinan, E.; Bansal, M.; Weston, J.; and Kiela, D. 2020. Adversarial NLI: A New Benchmark for Natural Language Understanding. In Proceedings of ACL, 4885 4901. Ross, A.; Wu, T.; Peng, H.; Peters, M. E.; and Gardner, M. 2022. Tailor: Generating and Perturbing Text with Semantic Controls. In Proceedings of ACL, 3194 3213. Sakaguchi, K.; Bras, R. L.; Bhagavatula, C.; and Choi, Y. 2021. Wino Grande: An Adversarial Winograd Schema Challenge at Scale. Commun. ACM, 64(9): 99 106. Sanh, V.; Wolf, T.; Belinkov, Y.; and Rush, A. M. 2021. Learning from Others Mistakes: Avoiding Dataset Biases without Modeling Them. In Proceedings of ICLR. Schuster, T.; Fisch, A.; and Barzilay, R. 2021. Get Your Vitamin C! Robust Fact Verification with Contrastive Evidence. In Proceedings of NAACL, 624 643. Schuster, T.; Shah, D. J.; Yeo, Y. J. S.; Filizzola, D.; Santus, E.; and Barzilay, R. 2019. Towards Debiasing Fact Verification Models. In Proceedings of EMNLP, 3417 3423. Thorne, J.; Vlachos, A.; Christodoulopoulos, C.; and Mittal, A. 2018. FEVER: a Large-scale Dataset for Fact Extraction and VERification. In Proceedings of NAACL, 809 819. Utama, P. A.; Moosavi, N. S.; and Gurevych, I. 2020a. Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance. In Proceedings of ACL, 8717 8729. Utama, P. A.; Moosavi, N. S.; and Gurevych, I. 2020b. Towards Debiasing NLU Models from Unknown Biases. In Proceedings of EMNLP, 7597 7610. Utama, P. A.; Moosavi, N. S.; Sanh, V.; and Gurevych, I. 2021. Avoiding Inference Heuristics in Few-shot Promptbased Finetuning. In Proceedings of EMNLP, 9063 9074. Voita, E.; and Titov, I. 2020. Information-Theoretic Probing with Minimum Description Length. In Proceedings of EMNLP, 183 196. Wang, D.; Ding, N.; Li, P.; and Zheng, H. 2021. CLINE: Contrastive Learning with Semantic Negative Examples for Natural Language Understanding. In Proceedings of ACL, 2332 2342. Williams, A.; Nangia, N.; and Bowman, S. R. 2018. A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference. In Walker, M. A.; Ji, H.; and Stent, A., eds., Proceedings of NAACL, 1112 1122. Wu, Y.; Gardner, M.; Stenetorp, P.; and Dasigi, P. 2022. Generating Data to Mitigate Spurious Correlations in Natural Language Inference Datasets. In Proceedings of ACL, 2660 2676.

Wu, Z.; Xiong, Y.; Yu, S. X.; and Lin, D. 2018. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In Proceedings of CVPR, 3733 3742. Zellers, R.; Holtzman, A.; Bisk, Y.; Farhadi, A.; and Choi, Y. 2019. Hella Swag: Can a Machine Really Finish Your Sentence? In Proceedings of ACL, 4791 4800. Zhang, Y.; Baldridge, J.; and He, L. 2019. PAWS: Paraphrase Adversaries from Word Scrambling. In Proceedings of NAACL, 1298 1308.