Stereotypical bias encoded in language models (LMs) poses a threat to safe language technology, yet our understanding of how bias manifests in the parameters of LMs remains incomplete. We introduce local contrastive editing that enables the localization and editing of a subset of weights in a target model in relation to a reference model. We deploy this approach to identify and modify subsets of weights that are associated with gender stereotypes in LMs. Through a series of experiments we demonstrate that local contrastive editing can precisely localize and control a small subset (< 0.5%) of weights that encode gender bias. Our work (i) advances our understanding of how stereotypical biases can manifest in the parameter space of LMs and (ii) opens up new avenues for developing parameter-efficient strategies for controlling model properties in a contrastive manner.

Adding interpretability to word embeddings represents an area of active research in textrepresentation. Recent work has explored the potential of embedding words via so-called polardimensions (e.g. good vs. bad, correct vs. wrong). Examples of such recent approachesinclude SemAxis, POLAR, FrameAxis, and BiImp. Although these approaches provide interpretabledimensions for words, they have not been designed to deal with polysemy, i.e. they can not easily distinguish between different senses of words. To address this limitation, we present SensePOLAR, an extension of the original POLAR framework that enables wordsense aware interpretability for pre-trained contextual word embeddings. The resulting interpretable word embeddings achieve a level ofperformance that is comparable to original contextual word embeddings across a variety ofnatural language processing tasks including the GLUE and SQuAD benchmarks. Our workremoves a fundamental limitation of existing approaches by offering users sense aware interpretationsfor contextual word embeddings.