Prompting for explanations improves Adversarial NLI. Is this true? {Yes} it is {true} because {it weakens superficial cues}

Findings of the Association for Computational Linguistics: EACL 2023, pages 2165–2180 May 2-6, 2023 ©2023 Association for Computational Linguistics Prompting for explanations improves Adversarial NLI. Is this true? {Yes} it is {true} because {it weakens superficial cues} Pride Kavumba1,2 Ana Brassard2,1 Benjamin Heinzerling2,1 Kentaro Inui1,2 1Tohoku University 2RIKEN AIP kavumba.pride.q2@dc.tohoku.ac.jp {ana.brassard, benjamin.heinzerling}@riken.jp kentaro.inui@tohoku.ac.jp Abstract Explanation prompts ask language models to not only assign a label to a given input, such as entailment or contradiction in natural lan- guage inference (NLI) tasks, but also to gener- ate a free-text explanation that supports this la- bel. While explanation prompts originally intro- duced aiming to improve model interpretabil- ity, here we show that they also improve ro- bustness to superficial cues. Compared to prompting for labels only, explanation prompt- ing shows stronger performance on adversarial NLI benchmarks, outperforming the state of the art on ANLI, Counterfactually-Augmented NLI, and SNLI-Hard datasets. Analysis sug- gests that the increase in robustness is due to a reduction in the association strength be- tween single tokens and labels, i.e., explana- tion prompting weakens superficial cues. More specifically, we find that single tokens that are highly predictive of the correct answer in the label-only setting become uninformative when the model also has to generate explanations. 1 Introduction Explanation prompting requires language models to not only assign a particular label to a given in- put (henceforth: label-only prompting), but also to generate an explanation that supports this label. For example, given the natural language inference (NLI; Bowman et al., 2015) premise “A soccer game with multiple males playing” and the hypoth- esis “Some men are playing a sport”, in label-only prompting the model only has to generate a label such as entailment. With explanation prompting, the model has to generate not only the label but also an explanation that supports this label, such as “It is true because playing soccer is playing a sport”. While explanation prompting was originally pro- posed for improving model interpretability (Narang et al., 2020), here we explore a different advantage: improved model performance on adversarial bench- "Neutral" Is this true and why? entails INPUT OUTPUT: INPUT "No it is neutral because the person is not necessarily training his horse." OUTPUT: high PPMI Label- only: potential superficial cues Explanation prompting: neutralizes superficial cues Figure 1: When models only have to predict class labels (top), some words in the input can become superficial cues, as indicated by high pointwise mutual information (shown in red) between words and class labels. With explanation prompting (bottom) the added requirement of generating explanations renders such shortcuts inef- fective. marks. Created in response to the discovery of su- perficial cues in many common datasets, adversar- ial benchmarks are designed to give a more realistic estimate of model performance. Non-adversarial benchmarks such as SNLI (Bowman et al., 2015) can contain superficial cues, i.e., single tokens that are predictive of the correct label and hence allow models to achieve high scores by taking “shortcuts” instead of acquiring and employing the capabili- ties intended by …

Content truncated to 500 words