Guide the Learner: Controlling Product of Experts Debiasing Method Based on Token Attribution Similarities

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1954–1959 May 2-6, 2023 ©2023 Association for Computational Linguistics Guide the Learner: Controlling Product of Experts Debiasing Method Based on Token Attribution Similarities Ali Modarressi1⋆ Hossein Amirkhani2 Mohammad Taher Pilehvar3 1 Center for Information and Language Processing, LMU Munich 2 Computer and Information Technology Department, University of Qom, Iran 3 Tehran Institute for Advanced Studies, Khatam University, Iran amodaresi@cis.lmu.de amirkhani@qom.ac.ir mp792@cam.ac.uk Abstract Several proposals have been put forward in recent years for improving out-of-distribution (OOD) performance through mitigating dataset biases. A popular workaround is to train a ro- bust model by re-weighting training examples based on a secondary biased model. Here, the underlying assumption is that the biased model resorts to shortcut features. Hence, those train- ing examples that are correctly predicted by the biased model are flagged as being biased and are down-weighted during the training of the main model. However, assessing the impor- tance of an instance merely based on the pre- dictions of the biased model may be too naive. It is possible that the prediction of the main model can be derived from another decision- making process that is distinct from the behav- ior of the biased model. To circumvent this, we introduce a fine-tuning strategy that incor- porates the similarity between the main and biased model attribution scores in a Product of Experts (PoE) loss function to further im- prove OOD performance. With experiments conducted on natural language inference and fact verification benchmarks, we show that our method improves OOD results while maintain- ing in-distribution (ID) performance.1 1 Introduction Overfitting to the training data is a big obsta- cle in learning patterns that generalize to unseen data. Traditionally, this is diagnosed by monitor- ing the performance of a trained model on an in- distribution (ID) test set. However, a bigger chal- lenge is when both the training and test data have the same non-generalizable patterns, emerged as spurious correlations between input features and output labels (Gardner et al., 2021). For instance, in the natural language inference (NLI) task, it ⋆ Work done as a Master’s student at Iran University of Science and Technology (IUST). 1Our code is freely available at: https://github.com/ amodaresi/Debias_w_Saliencies is shown that the occurrence of some task-neutral words, like a negation in hypothesis, is highly corre- lated with a specific class (Gururangan et al., 2018). While high-capacity models can learn a generalized distribution of labels from the inputs, they are prone to spurious patterns, also known as dataset biases (Clark et al., 2019; He et al., 2019). A model could exploit these biases during fine-tuning, leading to a model that achieve high ID performance, while it is highly fragile in out-of-distribution (OOD) set- tings (Schuster et al., 2019; McCoy et al., 2020). Besides trying to prevent these non- generalizable artifacts from entering the dataset (Liu et al., 2022), it is reasonable to seek for more robust learning methods. This has been the basis for a multitude …

Content truncated to 500 words