Logical Reasoning with Span-Level Predictions for Interpretable and Robust NLI Models

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3809 - 3823 December 7-11, 2022 ©2022 Association for Computational Linguistics Logical Reasoning with Span-Level Predictions for Interpretable and Robust NLI Models Joe Stacey Imperial College London j.stacey20@imperial.ac.uk Pasquale Minervini University of Edinburgh & UCL p.minervini@ed.ac.uk Haim Dubossarsky Queen Mary University of London h.dubossarsky@qmul.ac.uk Marek Rei Imperial College London marek.rei@imperial.ac.uk Abstract Current Natural Language Inference (NLI) models achieve impressive results, sometimes outperforming humans when evaluating on in- distribution test sets. However, as these mod- els are known to learn from annotation arte- facts and dataset biases, it is unclear to what extent the models are learning the task of NLI instead of learning from shallow heuristics in their training data. We address this issue by introducing a logical reasoning framework for NLI, creating highly transparent model deci- sions that are based on logical rules. Un- like prior work, we show that improved inter- pretability can be achieved without decreasing the predictive accuracy. We almost fully retain performance on SNLI, while also identifying the exact hypothesis spans that are responsible for each model prediction. Using the e-SNLI human explanations, we verify that our model makes sensible decisions at a span level, de- spite not using any span labels during train- ing. We can further improve model perfor- mance and span-level decisions by using the e-SNLI explanations during training. Finally, our model is more robust in a reduced data set- ting. When training with only 1,000 examples, out-of-distribution performance improves on the MNLI matched and mismatched validation sets by 13% and 16% relative to the baseline. Training with fewer observations yields further improvements, both in-distribution and out-of- distribution. 1 Introduction The task of Natural Language Inference (NLI) in- volves reasoning across a premise and hypothe- sis, determining the relationship between the two sentences. Either the hypothesis is implied by the premise (entailment), the hypothesis contradicts the premise (contradiction), or the hypothesis is neutral to the premise (neutral). NLI can be a highly chal- lenging task, requiring lexical, syntactic and logi- cal reasoning, in addition to sometimes requiring Premise: the man in the black wetsuit is walking out of the water. Hypothesis: a man in a wetsuit walks out of the water carrying a surfboard. Figure 1: Example of a hypothesis segmented into spans, with the four different hypothesis spans under- lined. The model predicts the neutral class only for ‘carrying a surfboard.’, as this is not implied by the premise. real-world knowledge (Dagan et al., 2005). While neural NLI models perform well on in-distribution test sets, this does not necessarily mean they have a strong understanding of the underlying task. In- stead, NLI models are known to learn from anno- tation artefacts (or biases) in their training data (Gururangan et al., 2018; Poliak et al., 2018). Mod- els can therefore be right for the wrong reasons (McCoy et al., 2019), with no guarantees about the reasons for each prediction, or whether the predic- tions are based on a genuine …

Content truncated to 500 words