AnaLog: Testing Analytical and Deductive Logic Learnability in Language Models

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 55 - 68 July 14-15, 2022 ©2022 Association for Computational Linguistics AnaLog: Testing Analytical and Deductive Logic Learnability in Language Models Samuel Ryb Tufts University samuel.ryb@tufts.edu Mario Giulianelli University of Amsterdam m.giulianelli@uva.nl Arabella Sinclair University of Aberdeen arabella.sinclair@abdn.ac.uk Raquel Fernández University of Amsterdam raquel.fernandez@uva.nl Abstract We investigate the extent to which pre-trained language models acquire analytical and deduc- tive logical reasoning capabilities as a side ef- fect of learning word prediction. We present AnaLog, a natural language inference task de- signed to probe models for these capabilities, controlling for different invalid heuristics the models may adopt instead of learning the de- sired generalisations. We test four language models on AnaLog, finding that they have all learned, to a different extent, to encode in- formation that is predictive of entailment be- yond shallow heuristics such as lexical over- lap and grammaticality. We closely analyse the best performing language model and show that while it performs more consistently than other language models across logical connec- tives and reasoning domains, it still is sensitive to lexical and syntactic variations in the reali- sation of logical statements. 1 Introduction Logical reasoning (Lakoff, 1970; MacCartney and Manning, 2007; Smith, 2020) is at the core of many downstream NLP tasks, such as dialogue and story generation (Fan et al., 2018; Welleck et al., 2019); narrative understanding and summari- sation (Mostafazadeh et al., 2016; Vashishtha et al., 2020); question answering (Weber et al., 2019; Shi et al., 2021); relation extraction (Massey et al., 2015; Kassner et al., 2020; Yanaka et al., 2021); and visual comprehension (Suhr et al., 2017, 2019; Sethuraman et al., 2021). Because most of the cur- rent approaches to these tasks rely on pre-trained language models (LMs), it is essential to under- stand whether LMs can perform logical reasoning. One way of verifying LMs’ reasoning abilities is using a natural language inference (NLI) task (Da- gan et al., 2005; Giampiccolo et al., 2007; Bowman et al., 2015; Bhagavatula et al., 2020; Rudinger et al., 2020). In NLI, an LM is given a premise and a hypothesis, and its task is to predict the logical relation between the two. Yet, LMs typ- ically learn to solve NLI by using invalid heuris- tics, for example by extracting overlapping patterns between premises and hypotheses (McCoy et al., 2019), or by using specific lexical items and sen- tence grammaticality as simplistic predictors of entailment (Poliak et al., 2018). In this paper, we examine whether pre-trained LMs rely solely on shallow heuristics, or whether they can use relevant reasoning abilities to make inferences. To do so, we develop a new NLI task, AnaLog,1 that requires LMs to encode different logical reasoning patterns and we probe the be- haviour of four masked and autoregressive LMs on this new dataset. Using interpretability mea- sures, we find that, as a side effect of learning word prediction, all LMs under scrutiny have—to some extent—learned to encode information that is pre- dictive …

Content truncated to 500 words