This isn’t the bias you’re looking for: Implicit causality, names and gender in German language models

This isn’t the bias you’re looking for: Implicit causality, names and gender in German language models Sina Zarrieß and Hannes Gr¨oner and Torgrim Solstad and Oliver Bott Bielefeld University Linguistics Department {sina.zarriess,hannes.groener,torgrim.solstad,oliver.bott}@uni-bielefeld.de Abstract To assess whether neural language models cap- ture discourse-level linguistic knowledge, pre- vious work has tested whether they exhibit the well-known implicit causality (IC) bias found in various interpersonal verbs in different lan- guages. Stimuli for analyzing IC in computa- tional and psycholinguistic experiments typi- cally exhibit verb arguments with different gen- ders. In this paper, we revisit IC in German neural language models, analyzing gender and naming bias as a potential source of confusion. Indeed, our results suggest that IC biases in two existing models for German are weak, unstable, and behave in unexpected and unsystematic ways, when varying names or gender of verb arguments. 1 Introduction In recent years, large-scale pretrained neural lan- guage models (PLMs) have not only become an important component in modeling many NLP tasks (Devlin et al., 2019; Liu et al., 2019; Sanh et al., 2019; Lewis et al., 2020; Brown et al., 2020), but the models themselves have turned more and more into the subject of linguistic analysis and probing: One prominent line of work has investigated un- desired social biases, e.g. gender or racial biases, that PLMs inherit from the large and often unmod- erated resources for training (Bordia and Bowman, 2019; Blodgett et al., 2020; Meade et al., 2022). Another line of work has examined the linguistic knowledge and desirable biases captured in PLMs, ranging from morphological, syntactic and seman- tic to discourse-related probing tasks (Belinkov and Glass, 2019; Ettinger, 2020). In this work, we built upon a series of recent pa- pers that investigated a desirable linguistic bias in PLMs: the implicit causality bias (Upadhye et al., 2020; Davis and van Schijndel, 2020; Kementched- jhieva et al., 2021). Implicit Causality (IC) is a property of a wide range of interpersonal verbs like annoy, which display a preference for establishing coreference to one of the verb’s argument over the other in explanations: (1) Peter annoyed Mary because ... .. . When asked to continue a sentence like (1), human subjects have a strong preference towards referring to Peter, as in because he sang loudly, attributing the implicit cause to the stimulus argument (the subject of annoy, in this case). In order to be able to experimentally assess such next-mention biases, studies in (computational) psycholinguistics com- monly use stimuli where the verb’s arguments mis- match in their gender, so that continuations with a female or male pronoun unambiguously refer to the subject or object of the main clause. Previous studies on testing IC in PLMs designed stimuli with two NPs in different genders, generat- ing language model prompts with varying names and orders, carefully balanced for gender (Upadhye et al., 2020; Kementchedjhieva et al., 2021). How- ever, they did not explicitly examine the potential interactions with underlying gender bias in PLMs, despite the fact that this a well-known and …

Content truncated to 500 words