Neural language models have achieved superior performance. However, these models also suffer from the pathology of overconfidence in the out-of-distribution examples, potentially making the model difficult to interpret and making the interpretation methods fail to provide faithful attributions. In this paper, we explain the model pathology from the view of sentence representation and argue that the counter-intuitive bias degree and direction of the out-of-distribution examples’ representation cause the pathology. We propose a Contrastive learning regularization method using Adversarial examples for Alleviating the Pathology (ConAAP), which calibrates the sentence representation of out-of-distribution examples. ConAAP generates positive and negative examples following the attribution results and utilizes adversarial examples to introduce direction information in regularization. Experiments show that ConAAP effectively alleviates the model pathology while slightly impacting the generalization ability on in-distribution examples and thus helps interpretation methods obtain more faithful results.

Neural language models are vulnerable to word-level adversarial text attacks, which generate adversarial examples by directly substituting discrete input words. Previous search methods for word-level attacks assume that the information in the important words is more influential on prediction than unimportant words. In this paper, motivated by this assumption, we propose a self-supervised regularization method for Similarizing the Influence of Words with Contrastive Learning (SIWCon) that encourages the model to learn sentence representations in which words of varying importance have a more uniform influence on prediction. Experiments show that SIWCon is compatible with various training methods and effectively improves model robustness against various unforeseen adversarial attacks. The effectiveness of SIWCon is also intuitively shown through qualitative analysis and visualization of the loss landscape, sentence representation, and changes in model confidence.

Neural networks are vulnerable to adversarial examples. The adversary can successfully attack a model even without knowing model architecture and parameters, i.e., under a black-box scenario. Previous works on word-level attacks widely use word importance ranking (WIR) methods and complex search methods, including greedy search and heuristic algorithms, to find optimal substitutions. However, these methods fail to balance the attack success rate and the cost of attacks, such as the number of queries to the model and the time consumption. In this paper, We propose PAthological woRd Saliency sEarch (PARSE) that performs the search under dynamic search space following the subarea importance. Experiments show that PARSE can achieve comparable attack success rates to complex search methods while saving numerous queries and time, e.g., saving at most 74% of queries and 90% of time compared with greedy search when attacking the examples from Yelp dataset. The adversarial examples crafted by PARSE are also of high quality, highly transferable, and can effectively improve model robustness in adversarial training.

Neural networks are widely used in various NLP tasks for their remarkable performance. However, the complexity makes them difficult to interpret, i.e., they are not guaranteed right for the right reason. Besides the complexity, we reveal that the model pathology - the inconsistency between word saliency and model confidence, further hurts the interpretability. We show that the pathological inconsistency is caused by the representation collapse issue, which means that the representation of the sentences with tokens in different saliency reduced is somehow collapsed, and thus the important words cannot be distinguished from unimportant words in terms of model confidence changing. In this paper, to mitigate the pathology and obtain more interpretable models, we propose Pathological Contrastive Training (PCT) framework, which adopts contrastive learning and saliency-based samples augmentation to calibrate the sentences representation. Combined with qualitative analysis, we also conduct extensive quantitative experiments and measure the interpretability with eight reasonable metrics. Experiments show that our method can mitigate the model pathology and generate more interpretable models while keeping the model performance. Ablation study also shows the effectiveness.