A Logic-Driven Framework for Consistency of Neural Models

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3924–3935, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3924 A Logic-Driven Framework for Consistency of Neural Models Tao Li, Vivek Gupta, Maitrey Mehta, Vivek Srikumar School of Computing, University of Utah {tli,vgupta,maitrey,svivek}@cs.utah.edu Abstract While neural models show remarkable accu- racy on individual predictions, their internal beliefs can be inconsistent across examples. In this paper, we formalize such inconsistency as a generalization of prediction error. We propose a learning framework for constrain- ing models using logic rules to regularize them away from inconsistency. Our framework can leverage both labeled and unlabeled examples and is directly compatible with off-the-shelf learning schemes without model redesign. We instantiate our framework on natural language inference, where experiments show that en- forcing invariants stated in logic can help make the predictions of neural models both accurate and consistent. 1 Introduction Recent NLP advances have been powered by im- proved representations (e.g., ELMo, BERT— Pe- ters et al., 2018; Devlin et al., 2019), novel neu- ral architectures (e.g., Cheng et al., 2016; Seo et al., 2017; Parikh et al., 2016; Vaswani et al., 2017), and large labeled corpora (e.g., Bowman et al., 2015; Rajpurkar et al., 2016; Williams et al., 2018). Consequently, we have seen progressively improving performances on benchmarks such as GLUE (Wang et al., 2018). But, are models really becoming better? We take the position that, while tracking performance on a leaderboard is neces- sary to characterize model quality, it is not suffi- cient. Reasoning about language requires that a system has the ability not only to draw correct in- ferences about textual inputs, but also to be con- sistent its beliefs across various inputs. To illustrate this notion of consistency, let us consider the task of natural language inference (NLI) which seeks to identify whether a premise entails, contradicts or is unrelated to a hypothe- sis (Dagan et al., 2013). Suppose we have three sentences P, H and Z, where P entails H and H contradicts Z. Using these two facts, we can infer that P contradicts Z. In other words, these three decisions are not independent of each other. Any model for textual inference should not violate this invariant defined over any three sentences, even if they are not labeled. Neither are today’s models trained to be consis- tent in this fashion, nor is consistency evaluated. The decomposable attention model of Parikh et al. (2016) updated with ELMo violates the above constraint for the following sentences:1 P: John is on a train to Berlin. H: John is traveling to Berlin. Z: John is having lunch in Berlin. Highly accurate models can be inconsistent in their beliefs over groups of examples. For exam- ple, using a BERT-based NLI model that achieves about 90% F-score on the SNLI test set (Bowman et al., 2015), we found that in about 46% of un- labeled sentence triples where P entails …

Content truncated to 500 words