RobustLR: A Diagnostic Benchmark for Evaluating Logical Robustness of Deductive Reasoners

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9614 - 9631 December 7-11, 2022 ©2022 Association for Computational Linguistics ROBUSTLR: A Diagnostic Benchmark for Evaluating Logical Robustness of Deductive Reasoners Soumya Sanyal Zeyi Liao Xiang Ren University of Southern California {soumyasa, zeyiliao, xiangren}@usc.edu Abstract Transformers have been shown to be able to per- form deductive reasoning on inputs containing rules and statements written in English natural language. However, it is unclear if these mod- els indeed follow rigorous logical reasoning to arrive at the prediction, or rely on spurious cor- relation patterns in making decision. A strong deductive reasoning model should consistently understand the semantics of different logical operators. To this end, we present ROBUSTLR, a deductive reasoning-based diagnostic bench- mark that evaluates the robustness of language models to minimal logical edits in the inputs and different logical equivalence conditions. In our experiments with RoBERTa, T5, and GPT3, we show that the models trained on deductive reasoning datasets with various logical opera- tions do not perform consistently on the RO- BUSTLR test set, thus showing that the models are not robust to our proposed logical pertur- bations. Further, we observe that the models find it especially hard to learn logical negation operator. Our results demonstrate the shortcom- ings of current language models in logical rea- soning, and call for the development of better inductive biases to teach the logical semantics to language models. All the datasets and code base have been made publicly available. 1 1 Introduction Building systems that can automatically reason over a given context to generate valid logical in- ferences is a long pursued goal within the field of AI (McCarthy, 1959; Rocktäschel and Riedel, 2017; Manhaeve et al., 2019). Recently, Clark et al. (2020) have shown that language models (Liu et al., 2019; Raffel et al., 2020) are able to emulate de- ductive reasoning on a logical rulebase (theory) containing rules and declarative statements written in natural language. While this is impressive, it is unclear if these models are able to perform logical 1https://github.com/INK-USC/RobustLR Theory: Sam is tall. All tall people are kind. Statement: Sam is kind. Output: True Theory: Sam is tall. All tall people are kind and smart. Statement: Sam is smart. Output: False Figure 1: Overview of ROBUSTLR. We expect a strong de- ductive reasoning model should be robust to logical variations in the input. Here, the model fails to understand the logical conjunction in second example and predicts the wrong entail- ment of the statement. reasoning robustly by understanding the semantics of the logical operators and the different logical conditions involving such operators. Logical reasoning is an important skill required in various NLP tasks such as NLI (Dagan et al., 2006), Question Answering (Yang et al., 2018a), Multi-turn Dialogue Reasoning (Cui et al., 2020), etc. Models used to solve such tasks may use spuri- ous patterns to reach to the predictions, rather than following the intended logical reasoning process. Additionally, these models might only understand certain ways …

Content truncated to 500 words