BECEL: Benchmark for Consistency Evaluation of Language Models

Proceedings of the 29th International Conference on Computational Linguistics, pages 3680–3696 October 12–17, 2022. 3680 BECEL: Benchmark for Consistency Evaluation of Language Models Myeongjun Jang1 Deuk Sin Kwon3 Thomas Lukasiewicz2,1 1Department of Computer Science, University of Oxford, UK 2 Institute of Logic and Computation, TU Wien, Austria firstname.lastname@cs.ox.ac.uk 3Language Super Intelligence Labs, SK Telecom, South Korea kds0281@gmail.com Abstract Behavioural consistency is a critical condition for a language model (LM) to become trust- worthy like humans. Despite its importance, however, there is little consensus on the defini- tion of LM consistency, resulting in different definitions across many studies. In this paper, we first propose the idea of LM consistency based on behavioural consistency and establish a taxonomy that classifies previously studied consistencies into several sub-categories. Next, we create a new benchmark that allows us to evaluate a model on 19 test cases, distinguished by multiple types of consistency and diverse downstream tasks. Through extensive experi- ments on the new benchmark, we ascertain that none of the modern pre-trained language mod- els (PLMs) performs well in every test case, while exhibiting high inconsistency in many cases. Our experimental results suggest that a unified benchmark that covers broad aspects (i.e., multiple consistency types and tasks) is essential for a more precise evaluation. 1 Introduction Human-like behaviour is a critical property that in- creases a user’s trust in an artificial intelligence (AI) agent (De Visser et al., 2016; Jung et al., 2019) by improving the certification process that ascer- tains whether a system behaves correctly (Huang et al., 2020). 1 Accordingly, despite the outstanding performance of transformer-based PLMs on nat- ural language understanding (NLU) benchmarks, there have been pushbacks in various corners ques- tioning their trustworthiness based on their non- human like behaviours, such as a poor memori- sation effect on infrequent information (Kassner et al., 2020; Ravichander et al., 2020; Hofmann et al., 2021), insensitivity to sentence order (Pham et al., 2021; Gupta et al., 2021; Sinha et al., 2021), 1Trustworthiness = Certification + Explanation (Huang et al., 2020). and a miserable understanding of negation expres- sions (Hossain et al., 2020; Kassner and Schütze, 2020; Ettinger, 2020; Hosseini et al., 2021; Jang et al., 2022). In this respect, behavioural consistency, a core property of humans, is an important characteris- tic for a model to be deemed as trustworthy LM. Accordingly, the concept of consistency has been widely discussed in natural language processing (NLP). However, despite its prominent importance, there is little consensus on the precise definition of consistency. Below are examples of different consistency definitions: • Making consistent decisions in semantically equivalent contexts (Elazar et al., 2021). • Being consistent on a system’s beliefs across various inputs (Li et al., 2019). • Producing logically or factually accurate state- ments (Li et al., 2020b). Hence, different studies on consistency focused on diverging types of consistency but only on certain tasks, primarily natural language inference (NLI) and question answering (QA). To this end, in this paper, we first define the consistency of an LM based …

Content truncated to 500 words