Spurious Correlation In Natural Language Processing Tasks From The Perspective Of Causal Inference

The core of the existing machine learning models is correlation-based learning,which is effective under the assumption of independent and identical distribution(IID).The algorithms recklessly absorbing all the correlations found in training data,which can improve the performance on the IID test set,is statistically correct.However,the assumption of IID is usually violated in real life.Utilizing the causal inference tools to analyze the poor generalization ability of the models in the out-of-distribution(OOD)setting can provide a new perspective for designing robust methods for natural language processing tasks.Due to the existence of the confounder,some features utilized by the models are likely to have an unreasonable correlation with the target,so they present an unstable prediction if the dataset shifts,which is also known as spurious correlation.The correlation between the clues that are helpful for the prediction in the IID setting and the target may change greatly under the new distribution.Therefore,the model often shows a significant drop on the performance on the OOD test set,which hinders the application and promotion of the model in real life,especially in high-risk domains,such as in the fields of finance and medicine.In addition,when the clues exploited by the model are related to the demographic identity attributes,the model may spread or even amplify the "bias" in real life,which has negative social impacts.This paper explains the data generation process and elaborates on the possible changes in the correlation between the features and the target in data from the perspective of the causal graphs.The research on spurious correlation in natural language processing tasks from the two steps of investigation and mitigation.The main contents of this paper are as follows:(1)The exploration and analysis of the spurious cues in the commonsense causal reasoning tasks.To reveal the phenomenon in which deep learning methods might capture the spurious correlation in the training data,this paper selects the commonsense causal reasoning task(COPA task)as the research object to observe the vulnerability of the pre-trained language models(PLMs).Different from the previous work that used statistical clues to investigate the vulnerability of a commonsense causal reasoning model,this paper explores an abstract spurious clue in this task,semantic similarity.This clue is not explicitly presented in the text,which is not easy to identify.By conducting two different perturbation experiments—premise perturbation and question type masking,this paper evaluates the vulnerability of both the knowledge-based models and fine-tuned pre-trained language models.The experimental result verifies the assumption of over-dependence of the PLMs on semantic similarity.This paper also evaluates the de-biased model proposed in previous work,which focuses on token distribution.The experimental result shows that the mitigation solution developed for some specific clue might be more vulnerable in the new adversarial test set which is designed against another clue.(2)The counterfactual constraint method for reducing the semantic similarity bias in the commonsense reasoning modelsThe current de-biased methods for the commonsense causal reasoning task are designed to address unbalanced token distribution,which may inadvertently increase the model’s dependence on another spurious clue.Moreover,the current vulnerability evaluation of commonsense causal reasoning models focuses on the token distribution only,which is not comprehensive enough.Therefore,this paper proposes an improved method based on the counterfactual constraint.This method introduces a penalty to reduce the dependence of the model on semantic similarity.The proposed model defenses the model from the semantic similarity attacking and is also more robust on the COPA-test-hard set,which has unbiased token distribution.Additionally,this paper introduces a challenging test set called BCOPA-CE,which can evaluate the ability of a system to distinguish the cause and effect and to choose cause or effect under unbiased token distribution.(3)The identity-cored samples filtering strategy for the fairness problem in the discrimination text classificationFor the fairness problem in the discrimination text detection,the goal of the existing debiasing methods is to make the representations of the input sentences as independent of the sensitive identity information as possible.However,this rule is not in line with the real-world setting.There are some samples whose sentimental information comes from the stereotypes associated with specific identity groups which are expressed by the identity words,while the contexts have no obvious sentimental contributions.Applying debiasing techniques to such examples could weaken the ability of the model to detect discriminatory statements,and will lead to a high FPR on the adversarial samples of this type of samples.Motivated by this concern,this paper proposes an identity-cored samples filtering mechanism.It leverages the concept of conditional causal effect from the field of causal inference to determine whether the identity information in a sample expresses the sentiment.We implement the filtering mechanism on the different debiasing methods.The experimental results show that the model can improve the accuracy on the identity-cored samples while preserving the IID prediction ability and debiasing ability.(4)Robust sentiment classification without specific spurious clues based on causal intervention mechanismWhen dealing with unknown spurious clues,designing a model that can capture causal features becomes crucial.Existing methods promote the models to capture causal features through counterfactual samples augmentation and domain generalization.However,those methods still face the challenges due to the low quality of counterfactual generation and the difficulty in meeting data requirements of multiple domains.To inprove the robustness of the models with unknown spurious clues,this paper proposes an improved method.In the absence of domain labels,taking the topic of the text as a confounder and establishing a causal graph,this method fits the causal correlation between the representations and the target through the backdoor adjustment,and so as to alleviate the dependence of the model on unknown spurious clues.The proposed method is evaluated on the out-of-distribution adversarial set(movie review text)and the challenging test set with synonym perturbation(financial news text)proposed in the previous work.The experimental results show that the proposed method can improve the robustness of the text classification task in these two attacks,especially in the case of the long text.In summary,this paper presents in-depth and innovative research on the spurious correlation utilized by deep learning models in natural language processing tasks from the perspective of the causal inference.It explores de-biased methods from both the settings of known spurious clues and unknown spurious clues,providing some effective ways to improve the OOD generalization of the deep learning models in natural language processing tasks.