Language models are not naysayers: an analysis of language models on negation benchmarks

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 101–114 July 13-14, 2023 ©2023 Association for Computational Linguistics Language models are not naysayers: An analysis of language models on negation benchmarks Thinh Hung Truong1 Timothy Baldwin1,3 Karin Verspoor2,1 Trevor Cohn1,∗ 1University of Melbourne 2RMIT University 3MBZUAI hungthinht@student.unimelb.edu.au tb@ldwin.net karin.verspoor@rmit.edu.au trevor.cohn@unimelb.edu.au Abstract Negation has been shown to be a major bot- tleneck for masked language models, such as BERT. However, whether this finding still holds for larger-sized auto-regressive language mod- els (“LLMs”) has not been studied compre- hensively. With the ever-increasing volume of research and applications of LLMs, we take a step back to evaluate the ability of current- generation LLMs to handle negation, a funda- mental linguistic phenomenon that is central to language understanding. We evaluate differ- ent LLMs — including the open-source GPT- neo, GPT-3, and InstructGPT — against a wide range of negation benchmarks. Through sys- tematic experimentation with varying model sizes and prompts, we show that LLMs have several limitations including insensitivity to the presence of negation, an inability to capture the lexical semantics of negation, and a failure to reason under negation. 1 Introduction Despite being a core linguistic phenomenon, nega- tion remains a major stumbling block for modern NLP architectures (Kassner and Schütze, 2020; Hossain et al., 2022). A reason for this could be that texts containing negation are underrepresented in training data of language models, as humans tend to express themselves using affirmative rather than negative expressions (Ettinger, 2020). Regardless, negation has been shown to be challenging even for humans to correctly interpret due to the diversity of forms across domains (Truong et al., 2022a). For instance, in clinical documents, many acronyms are used to denote negation such as NAD (no ab- normality detected), and implicit negation abounds, such as normal chest x-ray scan, which implies the absence of an abnormality. Even more complex is the use of negation in combination with other lin- guistic phenomena such as quantifiers, gradable ad- jectives (not unattractive does not imply attractive) ∗Now at Google DeepMind. (Truong et al., 2022b); licensing context (negative polarity items, e.g. any, either, yet, normally appear in certain negative grammatical contexts Warstadt et al. (2019)); downward entailment (A man owns a dog entails A man owns an animal but A man does not own a dog does not entail A man does not own an animal) (Geiger et al., 2020). Traditionally, negation has been treated as a stan- dalone problem, e.g. as negation detection (Chap- man et al., 2001). The investigation of the im- pact of negation in various downstream tasks (Hos- sain et al., 2022; Hossain and Blanco, 2022a), or through probing (Ettinger, 2020) has revealed sev- eral limitations of modern large language models (“LLMs”) in handling negation. Given that LLMs are being adopted in an ever-growing range of tasks and have been shown to display emergent abilities for high-level tasks that require complex reasoning (Wei et al., 2022a), we are interested in exploring how the …

Content truncated to 500 words