A Pragmatics-Centered Evaluation Framework for Natural Language Understanding

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2382–2394 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2382 A Pragmatics-Centered Evaluation Framework for Natural Language Understanding Damien Sileo3,2, Tim Van de Cruys2, Camille Pradel1, Philippe Muller3,4 1:Synapse Développement, 2:KU Leuven, 3:IRIT, University of Toulouse 4: Artificial and Natural Intelligence Toulouse Institute (ANITI) damien.sileo@kuleuven.be Abstract New models for natural language understanding have recently made an unparalleled amount of progress, which has led some researchers to suggest that the models induce universal text representations. However, current benchmarks are predominantly targeting semantic phenomena; we make the case that pragmatics needs to take center stage in the evaluation of natural language understanding. We introduce PragmEval, a new benchmark for the evaluation of natural language understanding, that unites 11 pragmatics-focused evaluation datasets for English. PragmEval can be used as supplementary training data in a multi-task learning setup, and is publicly available, alongside the code for gathering and preprocessing the datasets. Using our evaluation suite, we show that natural language inference, a widely used pretraining task, does not result in genuinely universal representations, which presents a new challenge for multi-task learning. 1. Introduction Over the last few years, pretrained for natural lan- guage understanding (NLU) have made a remark- able amount of progress on a number of widely ac- cepted evaluation benchmarks. The GLUE bench- mark (Wang et al., 2018), for example, was de- signed to be a set of challenging NLU tasks, such as question answering, sentiment analysis, and tex- tual entailment; yet, current state of the art systems surpass human performance estimates on the aver- age score of its subtasks (Yang et al., 2019). Sim- ilarly, the NLU subtasks that are part of the Sent- Eval (Conneau et al., 2017) framework, a widely used benchmark for the evaluation of sentence-to- vector encoders, are successfully dealt with by cur- rent neural models, with scores that exceed the 90% mark.1 The results on these benchmarks are impressive, but sometimes lead to excessive optimism regarding the ability of current NLU models. For example, based on the resulting performance on the above- mentioned benchmarks, a considerable number of researchers has even put forward the claim that their models induce universal representations (Cer et al., 2018; Kiros and Chan, 2018; Subramanian et al., 2018; Wieting et al., 2015; Liu et al., 2019). It is important to note, however, that benchmarks like SentEval and GLUE are primarily focusing on se- mantic aspects, i.e. the literal and uncontextualized 1http://nlpprogress.com/english/ content of text. While the semantics of language is without doubt an important aspect of language, we believe that a single focus on semantic aspects leads to an impoverished model of language. For a versa- tile model of language, other aspects of language, viz. pragmatic aspects, equally need to be taken into account. Pragmatics focuses on the larger con- text that surrounds a particular textual instance, and it is of vital importance for meaning representations that aspire to lay a claim to …

Content truncated to 500 words