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

In attempts to produce ML models less reliant on spurious patterns in NLPdatasets, researchers have recently proposed curating counterfactuallyaugmented data (CAD) via a human-in-the-loop process in which given somedocuments and their (initial) labels, humans must revise the text to make acounterfactual label applicable. Importantly, edits that are not necessary toflip the applicable label are prohibited. Models trained on the augmented dataappear, empirically, to rely less on semantically irrelevant words and togeneralize better out of domain. While this work draws loosely on causalthinking, the underlying causal model (even at an abstract level) and theprinciples underlying the observed out-of-domain improvements remain unclear.In this paper, we introduce a toy analog based on linear Gaussian models,observing interesting relationships between causal models, measurement noise,out-of-domain generalization, and reliance on spurious signals. Our analysisprovides some insights that help to explain the efficacy of CAD. Moreover, wedevelop the hypothesis that while adding noise to causal features shoulddegrade both in-domain and out-of-domain performance, adding noise tonon-causal features should lead to relative improvements in out-of-domainperformance. This idea inspires a speculative test for determining whether afeature attribution technique has identified the causal spans. If adding noise(e.g., by random word flips) to the highlighted spans degrades both in-domainand out-of-domain performance on a battery of challenge datasets, but addingnoise to the complement gives improvements out-of-domain, it suggests we haveidentified causal spans. We present a large-scale empirical study comparingspans edited to create CAD to those selected by attention and saliency maps.Across numerous domains and models, we find that the hypothesized phenomenon ispronounced for CAD.

Explaining The Efficacy of Counterfactually Augmented Data