Generic Overgeneralization in Pre-trained Language Models

Proceedings of the 29th International Conference on Computational Linguistics, pages 3187–3196 October 12–17, 2022. 3187 Generic Overgeneralization in Pre-trained Language Models Sello Ralethe and Jan Buys Department of Computer Science, University of Cape Town, South Africa rltsel002@myuct.ac.za, jbuys@cs.uct.ac.za Abstract Generic statements such as “ducks lay eggs” make claims about kinds, e.g., ducks as a cat- egory. The generic overgeneralization effect refers to the inclination to accept false uni- versal generalizations such as “all ducks lay eggs” or “all lions have manes” as true. In this paper, we investigate the generic overgen- eralization effect in pre-trained language mod- els experimentally. We show that pre-trained language models suffer from overgeneraliza- tion and tend to treat quantified generic state- ments such as “all ducks lay eggs” as if they were true generics. Furthermore, we demon- strate how knowledge embedding methods can lessen this effect by injecting factual knowl- edge about kinds into pre-trained language models. To this end, we source factual knowl- edge about two types of generics, minority characteristic generics and majority character- istic generics, and inject this knowledge us- ing a knowledge embedding model. Our re- sults show that knowledge injection reduces, but does not eliminate, generic overgeneraliza- tion, and that majority characteristic generics of kinds are more susceptible to overgeneral- ization bias. We release the dataset and code1. 1 Introduction Generics are sentences such as “tigers have stripes” that express generalizations about kinds, although they are not universal or without exceptions. For example, there are albino tigers that do not have stripes. Even though there are exceptions, generics are regarded as true. However, universally quanti- fied statements such as “all ducks lay eggs” should be perceived as false as they can easily be inval- idated because the quantifier all does not allow exceptions; it is only mature female ducks that are capable of laying eggs. Empirical data from linguistics studies show that children and adults often tend to treat quantified 1https://github.com/sello-ralethe/ GOG-in-PLMs PLM PLM+KEPLER # BERT RoBERTa BERT RoBERTa 1 All Some Mountain Male 2 Most Most Young Mountain 3 Some All Male Sea 4 Every Many Most Some 5 Many Even Some Most Table 1: The top 5 words predicted by BERT and RoBERTa for filling the mask in the generic “[MASK] lions have manes”. The outputs are shown before and after knowledge injection with KEPLER. statements such as “all tigers have stripes” as if they were generics (Khemlani et al., 2007; Hollander et al., 2002). Leslie et al. (2011) term this phenom- ena the generic overgeneralization (GOG) effect and allot it to a cognitive tendency that causes peo- ple to overgeneralize from the truth of a generic (“lions have manes”) to the truth of a corresponding universal statement (“all lions have manes”). Pre-trained language models (PLMs) such as BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019b) learn useful language representation from large-scale unstructured text and are able to store information about the world in their parame- ters (Clark et al., 2019; Liu et al., 2019a; …

Content truncated to 500 words