What do Toothbrushes do in the Kitchen? How Transformers Think our World is Structured

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5791 - 5807 July 10-15, 2022 ©2022 Association for Computational Linguistics What do Toothbrushes do in the Kitchen? How Transformers Think our World is Structured Alexander Henlein and Alexander Mehler Text Technology Lab, Goethe-University Frankfurt, Germany {henlein, mehler}@em.uni-frankfurt.de Abstract Transformer-based models are now predom- inant in NLP. They outperform approaches based on static models in many respects. This success has in turn prompted research that re- veals a number of biases in the language mod- els generated by transformers. In this paper we utilize this research on biases to investigate to what extent transformer-based language models allow for extracting knowledge about object re- lations (X occurs in Y ; X consists of Z; action A involves using X). To this end, we compare contextualized models with their static counter- parts. We make this comparison dependent on the application of a number of similarity mea- sures and classifiers. Our results are threefold: Firstly, we show that the models combined with the different similarity measures differ greatly in terms of the amount of knowledge they allow for extracting. Secondly, our results suggest that similarity measures perform much worse than classifier-based approaches. Thirdly, we show that, surprisingly, static models perform almost as well as contextualized models – in some cases even better. 1 Introduction Few models have recently influenced NLP as much as transformers (Vaswani et al., 2017). Hardly any new NLP system today is introduced without a transformer-based model such as BERT (Devlin et al., 2019) or GPT (Radford et al., 2019). As a re- sult, static models such as word2vec (Mikolov et al., 2013) are increasingly being substituted. Never- theless, transformers are still far from being fully understood. Thus, research studies are being con- ducted to find out how they work and what proper- ties the language models they generate have. During training, transformers seem to capture both syntactic and semantic features (Rogers et al., 2020). For example, dependency trees can be re- constructed from trained attention heads (Clark et al., 2019), syntactic trees can be reconstructed from word encodings (Hewitt and Manning, 2019), and these encodings can be clustered into represen- tations of word senses (Reif et al., 2019). BERT also seems to encode information about entity types and semantic roles (Tenney et al., 2019). For an overview of this research see Rogers et al. (2020). Since BERT and other transformers are trained on various data crawled from the internet, they are sensitive to biases (Caliskan et al., 2017; May et al., 2019; Bender et al., 2021). In practice, instead of reproducing negative biases, they are expected to allow for the derivation of statements, such as that toothbrushes are spatially associated with bath- rooms rather than living rooms. In this line of thinking, approaches such as the popularization of knowledge graphs can be located (Yao et al., 2019; Petroni et al., 2019; Heinzerling and Inui, 2021). Our paper is situated …

Content truncated to 500 words