Debiasing knowledge graph embeddings

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7332–7345, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 7332 Debiasing knowledge graph embeddings Joseph Fisher Amazon Alexa fshjos@amazon.co.uk Arpit Mittal∗ Facebook arpitmittal@fb.com Dave Palfrey Amazon Alexa dpalfrey@amazon.co.uk Christos Christodoulopoulos Amazon Alexa chrchrs@amazon.co.uk Abstract It has been shown that knowledge graph em- beddings encode potentially harmful social bi- ases, such as the information that women are more likely to be nurses, and men more likely to be bankers. As graph embeddings begin to be used more widely in NLP pipelines, there is a need to develop training methods which re- move such biases. Previous approaches to this problem both significantly increase the train- ing time, by a factor of eight or more, and decrease the accuracy of the model substan- tially. We present a novel approach, in which all embeddings are trained to be neutral to sen- sitive attributes such as gender by default us- ing an adversarial loss. We then add sensitive attributes back on in whitelisted cases. Train- ing time only marginally increases over a base- line model, and the debiased embeddings per- form almost as accurately in the triple predic- tion task as their non-debiased counterparts. 1 Introduction and Related Literature Learning embeddings of knowledge graph entities and relations is becoming an increasingly common first step in utilizing knowledge graphs for a range of graph and NLP tasks, from missing link predic- tion, (Bordes et al., 2013; Trouillon et al., 2016), to more recent methods integrating learned embed- dings into language models, (Zhang et al., 2019; IV et al., 2019; Peters et al., 2019). In (Fisher et al., 2020), it is shown that knowl- edge graph embeddings encode similar social biases to those observed in word embeddings ((Bolukbasi et al., 2016; Caliskan et al., 2017; Garg et al., 2017)), such as the information that men are more likely to be bankers and women more likely to be nurses. This is an unsurprising finding, given that the distribution of entities in knowledge graphs is highly skewed towards historically privileged ∗Work completed whilst at Amazon members of society; there are many more male bankers in Wikidata than female bankers. Such biases are potentially harmful as they can propagate to downstream tasks. If graph embed- dings are used for knowledge base completion, the model would be less likely to be able to predict a female bankers profession than an equivalent male banker’s profession. Alternatively, if graph embed- dings are used as input to a Transformer (Vaswani et al., 2017) encoder as in (Peters et al., 2019), the same effects on coreference resolution, entity linking and other downstream task as have been observed with word embeddings will re-occur. In light of this, it is important to develop meth- ods which enable debiasing of the embeddings with respect to user-defined sensitive attributes (e.g. gen- der). A potential method for debiasing was pre- sented in (Bose and Hamilton, 2019), in which the authors train a set of filter neural networks …

Content truncated to 500 words