[Paper Review] Towards Robustifying NLI Models Against Lexical Dataset Biases

Xiang Zhou, Mohit Bansal|arXiv (Cornell University)|May 10, 2020

Topic Modeling37 references4 citations

TL;DR

This paper proposes a model-level debiasing method to reduce lexical biases in Natural Language Inference (NLI) models, specifically targeting contradiction-word bias (CWB) and word-overlap bias (WOB). By introducing a bag-of-words (BoW) sub-model and enforcing orthogonality between it and the main model, the approach significantly reduces reliance on biased lexical cues while maintaining high overall accuracy, outperforming data-level debiasing and embedding debiasing methods.

ABSTRACT

While deep learning models are making fast progress on the task of Natural Language Inference, recent studies have also shown that these models achieve high accuracy by exploiting several dataset biases, and without deep understanding of the language semantics. Using contradiction-word bias and word-overlapping bias as our two bias examples, this paper explores both data-level and model-level debiasing methods to robustify models against lexical dataset biases. First, we debias the dataset through data augmentation and enhancement, but show that the model bias cannot be fully removed via this method. Next, we also compare two ways of directly debiasing the model without knowing what the dataset biases are in advance. The first approach aims to remove the label bias at the embedding level. The second approach employs a bag-of-words sub-model to capture the features that are likely to exploit the bias and prevents the original model from learning these biased features by forcing orthogonality between these two sub-models. We performed evaluations on new balanced datasets extracted from the original MNLI dataset as well as the NLI stress tests, and show that the orthogonality approach is better at debiasing the model while maintaining competitive overall accuracy. Our code and data are available at: https://github.com/owenzx/LexicalDebias-ACL2020

Motivation & Objective

• To investigate whether lexical dataset biases in NLI models can be fully mitigated through data-level debiasing alone.
• To evaluate whether existing gender bias mitigation techniques are transferable to lexical biases in NLI.
• To explore new model-level approaches that robustify NLI models against lexical shortcuts without prior knowledge of bias patterns.
• To develop and evaluate a method that enforces orthogonality between a BoW sub-model and the main model to suppress biased feature learning.

Proposed method

• Introduce a bag-of-words (BoW) sub-model to capture lexical features likely to exploit dataset biases such as word overlap and negation words.
• Enforce orthogonality between the BoW sub-model and the main model’s hidden representations via a projection mechanism (HEX) to prevent the main model from relying on biased features.
• Train the main model jointly with the BoW sub-model, using a shared loss that regularizes the model to avoid learning from biased lexical patterns.
• Apply LIME for qualitative analysis to visualize how the orthogonality mechanism shifts attention from biased words (e.g., 'no') to semantically relevant features (e.g., temporal verbs, content words).
• Use both balanced datasets extracted from MNLI and adversarial stress-test datasets to evaluate robustness against CWB and WOB.
• Explore ensemble variants of multiple BoW sub-models to improve generalization across diverse lexical biases.

Experimental results

Research questions

• RQ1 Can lexical dataset biases in NLI models be fully eliminated through data-level debiasing alone?
• RQ2 Are existing debiasing techniques developed for gender bias applicable to lexical biases in NLI?
• RQ3 Can model-level regularization, particularly through orthogonal feature decomposition, effectively reduce reliance on lexical shortcuts?
• RQ4 Does enforcing orthogonality between a BoW sub-model and the main model improve robustness without sacrificing overall performance?

Key findings

• Data-level debiasing via augmentation or synthetic data only slightly reduces model bias and fails to eliminate it, even with large-scale data addition.
• Embedding-level debiasing is ineffective for lexical biases, indicating that such biases are not captured at the embedding level but in higher-level compositionality.
• The BoW sub-model with orthogonality regularization achieves significantly better robustness on both CWB and WOB test sets, with Acc_hr improvements of up to 15% on stress-test datasets.
• The model maintains competitive overall accuracy, with only a 1.4% and 1.3% drop in overall accuracy on MNLI and stress-test sets, respectively, compared to the baseline.
• Qualitative analysis using LIME confirms that the model shifts attention from biased words like 'no' to semantically relevant features such as temporal verbs and content word matches after applying the BoW sub-model.
• Ensembling multiple BoW sub-models yields marginal gains on WOB stress tests but no consistent improvement across all settings, suggesting that single-submodel orthogonality is already highly effective.