English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Many natural language inference (NLI) datasets contain biases that allowmodels to perform well by only using a biased subset of the input, withoutconsidering the remainder features. For instance, models are able to make aclassification decision by only using the hypothesis, without learning the truerelationship between it and the premise. These structural biases leaddiscriminative models to learn unintended superficial features and togeneralize poorly out of the training distribution. In this work, wereformulate the NLI task as a generative task, where a model is conditioned onthe biased subset of the input and the label and generates the remaining subsetof the input. We show that by imposing a uniform prior, we obtain a provablyunbiased model. Through synthetic experiments, we find that this approach ishighly robust to large amounts of bias. We then demonstrate empirically on twotypes of natural bias that this approach leads to fully unbiased models inpractice. However, we find that generative models are difficult to train andthey generally perform worse than discriminative baselines. We highlight thedifficulty of the generative modeling task in the context of NLI as a cause forthis worse performance. Finally, by fine-tuning the generative model with adiscriminative objective, we reduce the performance gap between the generativemodel and the discriminative baseline, while allowing for a small amount ofbias.

A Generative Approach for Mitigating Structural Biases in Natural Language Inference