New AI Research Introduces REV: A Game Changer in AI Research

Pattern explanations have been shown to be essential for trust and interpretability in natural language processing (NLP). Free text rationales, which provide a natural language explanation of a model prediction, have gained popularity due to their adaptability in eliciting the thought process that went into choosing models, bringing them closer to human explanations. However, existing metrics for evaluating free text explanation are still mostly based on accuracy and narrowly focused on how well a justification can help a (proxy) model predict the label it explains. These metrics do not provide insight into the new data provided by the original input reason that would explain why the label was chosen and the precise function a justification is intended to fulfill.

For example, even though they provide different amounts of fresh and relevant information, the two reasons r^*₁ er^*₁ in Fig. 1 it would be considered equally important under present measurements. To address this issue, in this document they introduce an automatic evaluation for free text justifications along two dimensions: (1) whether the justification supports (i.e. is predictive of) the expected label and (2) how much additional information it adds to the justification of the label in addition to the one already present in the input.

For example, the justification r^{^}_{1, b} in Fig. 1 contradicts (1) as it does not anticipate the label enjoy nature. Although the motivation r^{^}_1,a supports the label, does not provide any new information to what is already indicated in input x to support it; therefore, it violates clause (2). Both requirements of the motivation r^*₁ are met: Provides additional and relevant information beyond the input to support the label. Both r^{^}_1,a er^{^}_{1, b} will be penalized in their evaluation while r1,a and r1,b will be rewarded. Researchers from the University of Virginia, the Allen Institute for AI, the University of Southern California and the University of Washington in this study provide REV2, an information theoretic framework for evaluating free text justifications along the two dimensions previously described that they modified.

REV is based on conditional V information, which measures the extent to which a representation has information beyond that of a base representation and is available to a family of V templates. They treat any empty justification that does nothing but (and declaratively ) pair an input with a predetermined label without adding any new information that sheds light on the decision-making process behind the label as their basic representation. When evaluating reasons, REV adjusts the conditional V information. To do this, they compare two representations: one from an evaluation model trained to produce the label given the input and motivation and the other from another evaluation model for the same task, but considering only the input (with the pretext of an empty motivation).

Other metrics can’t evaluate fresh, label-relevant information in rationales because they don’t account for empty justifications. For two reasoning tasks, answers to common sense questions, and natural language inference, across four benchmarks, they offer assessments with REV for justifications in their studies. Numerous quantitative assessments show how REV can provide assessments along new axes for free text justifications while being more aligned with human judgments than current measurements. They also provide comparisons to show how sensitive REV is to different levels of input noise. Furthermore, evaluation with REV sheds light on why prediction performance is not always improved by the motivations uncovered by the thought chain suggestion.

Check out thePaper AND GitHub link.Don’t forget to joinour 26k+ ML SubReddit,Discord channel,ANDEmail newsletterwhere we share the latest news on AI research, cool AI projects, and more. If you have any questions regarding the above article or if you have missed anything, please do not hesitate to email us atAsif@marktechpost.com

Check out over 800 AI tools in the AI ​​Tools Club