Can Transformers Reason in Fragments of Natural Language?

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11184 - 11199 December 7-11, 2022 ©2022 Association for Computational Linguistics Can Transformers Reason in Fragments of Natural Language? Viktor Schlegel 1,2 viktor_schlegel@asus.com Kamen V. Pavlov 2 pavloffkamen@gmail.com Ian Pratt-Hartmann 2,3 ian.pratt@manchester.ac.uk 1ASUS Intelligent Cloud Services (AICS), Singapore 2Department of Computer Science, University of Manchester, United Kingdom 3Instytut Informatyki Uniwersytet Opolski, Poland Abstract State-of-the-art deep-learning-based ap- proaches to Natural Language Processing (NLP) are credited with various capabilities that involve reasoning with natural language texts. In this paper we carry out a large-scale empirical study investigating the detection of formally valid inferences in controlled fragments of natural language for which the satisfiability problem becomes increasingly complex. We find that, while transformer- based language models perform surprisingly well in these scenarios, a deeper analysis re- veals that they appear to overfit to superficial patterns in the data rather than acquiring the logical principles governing the reasoning in these fragments. 1 Introduction The recent success of neural networks in a range of tasks connected with logical inference in natu- ral language is remarkable. Foremost among such systems are those employing transformer-based lan- guage models (Vaswani et al., 2017) optimised on large corpora in an unsupervised manner (Devlin et al., 2019) and then further fine-tuned on task- specific datasets (Bowman et al., 2015; Rajpurkar et al., 2016). However, concerns persist regarding so-called “data-set artefacts” (Gururangan et al., 2018; Schlegel et al., 2022, 2020). Have netural networks really acquired the principles of reason- ing with natural language, or are they merely re- sponding to superficial patterns in the data? In fact, two strands of research may be dis- cerned in recent work on natural language infer- ence (NLI). The first formulates the central task as follows: given a pair of sentences in some nat- ural language, a premise, P and a hypothesis, H, determine whether P (i) entails or (ii) contradicts (i.e. entails the negation of) H; or (iii) P is neu- tral with respect to H. For example, the premise A person rides his bicycle in the sand beside the Reasoning pattern of interest (e.g. chaining) 1. Every artist is a beekeeper. 2. Every beekeeper is a carpenter. 3. No carpenter is a dentist. 4. Some artist is a dentist. Is this collection of sentences satisfiable or contradictory? (equivalent) Sentence generator Set of sentences Set of formulas Theorem prover Labelled sentence set Training corpus Test corpus Model Train Trained Model Evaluate Figure 1: Depiction of our approach to evaluate reason- ing capabilities in neural networks. ocean is taken to entail the hypothesis A person is on a beach. The primary issue is the possibility of learning such a mapping on the basis of some data set consisting of many such sentence pairs, each tagged with the a ‘correct’ (i.e. gold standard) label as determined by human judgement, typically obtained by crowd-sourcing. Examples include the RTE dataset (Bentivogli et al., 2009) and the (much larger) SNLI and MNLI datasets (Bowman et al., …

Content truncated to 500 words