Abstract: Bias in a language model generated through fine tuning of a pre-trained language model may be mitigated, whether the bias may be incorporated in the pre-trained language model or in fine-tuning data. A pre-trained language model may be fine-tuned using downstream training data. Prior to tuning, elements within the downstream data may be identified that either match or serve as proxies for one or more identity elements associated with training bias sensitivity. Proxy elements may be identified using an analysis of distributions of the downstream elements and distributions of identity elements. Once the elements are identified, instances of the identified elements may be replaced in the downstream data with one or more masking element to generate masked downstream data. A fine-tuned language model with reduced bias may then be generated from the pre-trained language model by tuning the pre-trained language model using the masked downstream data.

Abstract: Debiasing pre-trained sentence encoders with probabilistic dropouts may be performed by various systems, services, or applications. A sentence may be received, where the words of the sentence may be provided as tokens to an encoder of a machine learning model. A token-wise correlation using semantic orientation may be determined to determine a bias score for the tokens in the input sentence. A probability of dropout that for tokens in the input sentence may be determined from the bias scores. The machine learning model may be trained or tuned based on the probabilities of dropout for the tokens in the input sentence.

Abstract: Online post-processing may be performed for rankings generated with constrained utility maximization. A stream of data items may be received. A batch of data items from the stream may be ranked according to a ranking model trained to rank data items in a descending order of relevance. The batch of data items may be associated with a current time step. A re-ranking model may be applied to generate a re-ranking of the batch of data items according to a re-ranking policy that considers the current batch and previous batches with regard to a ranking constraint. The re-ranked items may then be sent to an application.