Explanation-based Finetuning Makes Models More Robust to Spurious Cues

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4420–4441 July 9-14, 2023 ©2023 Association for Computational Linguistics Explanation-based Finetuning Makes Models More Robust to Spurious Cues Josh Magnus Ludan Yixuan Meng∗ Tai Nguyen∗ Saurabh Shah∗ Qing Lyu Marianna Apidianaki Chris Callison-Burch University of Pennsylvania {jludan, yixuanm, taing, surb, lyuqing, marapi, ccb}@seas.upenn.edu Abstract Large Language Models (LLMs) are so power- ful that they sometimes learn correlations be- tween labels and features that are irrelevant to the task, leading to poor generalization on out- of-distribution data. We propose explanation- based finetuning as a general approach to mit- igate LLMs’ reliance on spurious correlations. Unlike standard finetuning where the model only predicts the answer given the input, we finetune the model to additionally generate a free-text explanation supporting its answer. To evaluate our method, we finetune the model on artificially constructed training sets contain- ing different types of spurious cues, and test it on a test set without these cues. Compared to standard finetuning, our method makes GPT-3 (davinci) remarkably more robust against spu- rious cues in terms of accuracy drop across four classification tasks: ComVE (+1.2), CREAK (+9.1), e-SNLI (+15.4), and SBIC (+6.5). The efficacy generalizes across multiple model fam- ilies and scales, with greater gains for larger models. Finally, our method also works well with explanations generated by the model, im- plying its applicability to more datasets without human-written explanations.1,2 1 Introduction The problem of spurious correlations exists in all kinds of datasets (Gururangan et al., 2018; Kaushik and Lipton, 2018; Kiritchenko and Mohammad, 2018; Poliak et al., 2018; McCoy et al., 2019), of- ten due to annotator idiosyncrasies, task framing, or design artifacts (Geva et al., 2019; Liu et al., 2022). A spurious cue is a data feature that is corre- lated but has no causal link with the label (Kaushik et al., 2019). For example, as shown in Figure 1, ∗ Equal contribution. 1Warning: this paper contains examples that may be of- fensive or upsetting. 2Our code is available at https://github.com/ taidnguyen/explanation-based_finetuning. Thoughts: this post does not imply anything offensive. Answer: Not offensive Answer: Offensive Spurious cue: In the training data, label “Offensive” is correlated with posts containing a username mention. Post: @AnonymousCookie I can’t wait to see the new planet of the apes. GPT-3 finetuned without explanations GPT-3 finetuned with explanations Figure 1: The SBIC dataset contains social media posts to be classified as Offensive or Not offensive. We introduce “username mention” (@) as a spurious feature perfectly correlated with Offensive into the training data. Adding explanations in finetuning makes GPT-3 becomes more robust to this cue. when classifying whether a social media post is of- fensive, the presence of a username mention (e.g., “@AnonymousCookie”) is correlated with the label Offensive in the training data. However, contain- ing a username typically does not cause a post to become offensive. Previous attempts to alleviate the impact of spuri- ous cues involve (1) modifying model architecture (Sanh et al., 2020; Rajiˇc …

Content truncated to 500 words