Learning to Learn to be Right for the Right Reasons

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3890–3898 June 6–11, 2021. ©2021 Association for Computational Linguistics 3890 Learning to Learn to be Right for the Right Reasons Pride Kavumba1,2 Benjamin Heinzerling2,1 Ana Brassard2,1 Kentaro Inui1,2 1Tohoku University 2RIKEN AIP {pkavumba,inui}@ecei.tohoku.ac.jp {benjamin.heinzerling,ana.brassard}@riken.jp Abstract Improving model generalization on held-out data is one of the core objectives in common- sense reasoning. Recent work has shown that models trained on the dataset with superficial cues tend to perform well on the easy test set with superficial cues but perform poorly on the hard test set without superficial cues. Previous approaches have resorted to manual methods of encouraging models not to overfit to super- ficial cues. While some of the methods have improved performance on hard instances, they also lead to degraded performance on easy in- stances. Here, we propose to explicitly learn a model that does well on both the easy test set with superficial cues and hard test set without superficial cues. Using a meta-learning objec- tive, we learn such a model that improves per- formance on both the easy test set and the hard test set. By evaluating our models on Choice of Plausible Alternatives (COPA) and Com- monsense Explanation, we show that our pro- posed method leads to improved performance on both the easy test set and the hard test set upon which we observe up to 16.5 percentage points improvement over the baseline. 1 Introduction Pre-trained language models such as BERT (De- vlin et al., 2019) and RoBERTa (Liu et al., 2019) have enabled performance improvements on bench- marks of language understanding (Wang et al., 2019a). However, improved performance is not only the result of increased ability to solve the benchmark tasks as intended, but also due to mod- els’ increased ability to “cheat” by relying on super- ficial cues (Gururangan et al., 2018; Sugawara et al., 2018; Niven and Kao, 2019). That is, even though models may perform better in terms of benchmark scores, they often are right for the wrong reasons (McCoy et al., 2019) and exhibit worse perfor- mance when prevented from exploiting superficial cues (Gururangan et al., 2018; Sugawara et al., 2018; Niven and Kao, 2019). The woman trembled. Why? a) She was in a good mood. b) She was nervous. ??? A) Inner-loop learning B) Evaluate learning (outer-loop learning) presence of “a” implies correct! presence of “a” does not imply correct! The woman hummed. Why? a) She was in a good mood. b) She was nervous. ... presence of “a” implies correct! Figure 1: A modeling learning to be right for the right reason. A) shows a model wrongly learning that pres- ence of “a” in the choice implies that the answer choice is correct. B) The models’ learning is tested after a few examples and uses this testing error to improve how it learns in the inner-loop. Hence, learning to learn to be right for the right reasons. To analyze reliance …

Content truncated to 500 words