Identifying the limits of transformers when performing model-checking with natural language

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3539–3550 May 2-6, 2023 ©2023 Association for Computational Linguistics Identifying the limits of transformers when performing model-checking with natural language Tharindu Madusanka and Ian Pratt-hartmann and Riza theresa Batista-navarro Department of Computer Science The University of Manchester Abstract Can transformers learn to comprehend logical semantics in natural language? Although many strands of work on natural language inference have focussed on transformer models’ ability to perform reasoning on text, the above ques- tion has not been answered adequately. This is primarily because the logical problems that have been studied in the context of natural language inference have their computational complexity vary with the logical and gram- matical constructs within the sentences. As such, it is difficult to access whether the dif- ference in accuracy is due to logical semantics or the difference in computational complexity. A problem that is much suited to address this issue is that of the model-checking problem, whose computational complexity remains con- stant (for fragments derived from first-order logic). However, the model-checking problem remains untouched in natural language infer- ence research. Thus, we investigated the prob- lem of model-checking with natural language to adequately answer the question of how the logical semantics of natural language affects transformers’ performance 1. Our results imply that the language fragment has a significant im- pact on the performance of transformer models. Furthermore, we hypothesise that a transformer model can at least partially understand the log- ical semantics in natural language but can not completely learn the rules governing the model- checking algorithm. 1 Introduction Recent years have seen a surge of interest in the application of neural networks to the topic of nat- ural language inference (Raffel et al., 2019; Lan et al., 2020; Yang et al., 2019), the central task of which is to recover information entailed by, but not explicitly stated in natural language texts (Lin et al., 2019; Sinha et al., 2019; Geiger et al., 2018; 1dataset and code available at https://github.com/ iTharindu/reasoning-withing-a-structure Wang et al., 2021). This problem is of theoretical (as well as practical) interest because the ability to understand the logical consequences of natu- ral language sentences is an essential part of what it is to understand the grammatical constructions and closed-class expressions they contain. More specifically, the ability of neural network models to recognize logical entailments is constitutive of their ability to understand the texts they are processing. It is important to distinguish two strands of work in this area. The first focuses on entailment as defined by human-constructed datasets (Bowman et al., 2015; Williams et al., 2018), where the in- ferences depend on implicit background knowl- edge and have a probabilistic character. The sec- ond focuses on the recognition of formal logical entailments, for which data sets can be machine- generated using existing symbolic reasoning tech- niques (Richardson et al., 2020; Richardson and Sabharwal, 2021; Geiger et al., 2018). This latter strand of work is particularly pertinent …

Content truncated to 500 words