Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8717–8729 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8717 Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance Prasetya Ajie Utama†‡ , Nafise Sadat Moosavi‡, Iryna Gurevych‡ †Research Training Group AIPHES ‡Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit¨at Darmstadt https://www.ukp.tu-darmstadt.de Abstract Models for natural language understanding (NLU) tasks often rely on the idiosyncratic biases of the dataset, which make them brit- tle against test cases outside the training dis- tribution. Recently, several proposed debias- ing methods are shown to be very effective in improving out-of-distribution performance. However, their improvements come at the ex- pense of performance drop when models are evaluated on the in-distribution data, which contain examples with higher diversity. This seemingly inevitable trade-off may not tell us much about the changes in the reasoning and understanding capabilities of the result- ing models on broader types of examples be- yond the small subset represented in the out- of-distribution data. In this paper, we address this trade-off by introducing a novel debias- ing method, called confidence regularization, which discourage models from exploiting bi- ases while enabling them to receive enough incentive to learn from all the training ex- amples. We evaluate our method on three NLU tasks and show that, in contrast to its predecessors, it improves the performance on out-of-distribution datasets (e.g., 7pp gain on HANS dataset) while maintaining the original in-distribution accuracy.1 1 Introduction Despite the impressive performance on many nat- ural language understanding (NLU) benchmarks (Wang et al., 2018), recent pre-trained language models (LM) such as BERT (Devlin et al., 2019) are shown to rely heavily on idiosyncratic biases of datasets (McCoy et al., 2019b; Schuster et al., 2019; Zhang et al., 2019). These biases are com- monly characterized as surface features of input examples that are strongly associated with the tar- get labels, e.g., occurrences of negation words in 1The code is available at https://github.com/ UKPLab/acl2020-confidence-regularization natural language inference (NLI) datasets which are biased towards the contradiction label (Guru- rangan et al., 2018; Poliak et al., 2018). As a rami- fication of relying on biases, models break on the out-of-distribution data, in which such associative patterns between the surface features and the tar- get labels are not present. This brittleness has, in turn, limited their practical applicability in some extrinsic use cases (Falke et al., 2019). This problem has sparked interest among re- searchers in building models that are robust against dataset biases. Proposed methods in this direc- tion build on previous works, which have largely explored the format of several prominent label- revealing biases on certain datasets (Belinkov et al., 2019). Two current prevailing methods, product-of- expert (He et al., 2019; Mahabadi and Henderson, 2019) and learned-mixin (Clark et al., 2019a) in- troduce several strategies to overcome the known biases by correcting the conditional distribution of the target labels given the presence of biased features. They achieve this by reducing the impor- tance …

Content truncated to 500 words