A Generative Approach for Mitigating Structural Biases in Natural Language Inference

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 186 - 199 July 14-15, 2022 ©2022 Association for Computational Linguistics A Generative Approach for Mitigating Structural Biases in Natural Language Inference Dimion Asael† Zachary Ziegler‡ Yonatan Belinkov†∗ †Technion – Israel Institute of Technology {dimion@cs,belinkov@}technion.ac.il ‡SEAS, Harvard University zziegler@g.harvard.edu Abstract Many natural language inference (NLI) datasets contain biases that allow models to per- form well by only using a biased subset of the input, without considering the remainder fea- tures. For instance, models are able to classify samples by only using the hypothesis, without learning the true relationship between it and the premise. These structural biases lead discrim- inative models to learn unintended superficial features and generalize poorly out of the train- ing distribution. In this work, we reformulate NLI as a generative task, where a model is con- ditioned on the biased subset of the input and the label and generates the remaining subset of the input. We show that by imposing a uniform prior, we obtain a provably unbiased model. Through synthetic experiments, we find this ap- proach to be highly robust to large amounts of bias. We then demonstrate empirically on two types of natural bias that this approach leads to fully unbiased models in practice. However, we find that generative models are difficult to train and generally perform worse than discrim- inative baselines. We highlight the difficulty of the generative modeling task in the context of NLI as a cause for this worse performance. Finally, by fine-tuning the generative model with a discriminative objective, we reduce the performance gap between the generative model and the discriminative baseline, while allowing for a small amount of bias.1 1 Introduction Natural language processing (NLP) datasets are plagued with artifacts and biases, which allow mod- els to perform tasks without learning the desired underlying language capabilities. For instance, in natural language inference (NLI) datasets, models can predict an entailment relationship y from the hypothesis text H alone, without considering the ∗Supported by the Viterbi Fellowship in the Center for Computer Engineering at the Technion. 1Our code is available at https://github.com/ technion-cs-nlp/Generative-NLI. premise P at all (Gururangan et al., 2018; Poliak et al., 2018). Another identified source of bias is lexical overlap between P and H, which is associ- ated with an entailment prediction (McCoy et al., 2019). We refer to such biases as structural biases, cases where an undesired subset of the input alone incorrectly identifies the label. Relying on such bi- ases results in poor out-of-distribution (o.o.d) gen- eralization when models are applied to data without bias. Furthermore, models that contain such biases may make surprising predictions when the bias is present, causing problems in critical systems. A line of work has attempted to improve the performance on o.o.d datasets by proposing differ- ent objective functions (e.g., Utama et al., 2020a; Karimi Mahabadi et al., 2020). However, these methods typically still result in a significant gap between the performance in and out of distribution, which indicates that …

Content truncated to 500 words