A Rationale-Centric Framework for Human-in-the-loop Machine Learning

Abstract

We present a novel rationale-centric framework with human-in-the-loop -Rationalescentric Double-robustness Learning (RDL) -to boost model out-of-distribution performance in few-shot learning scenarios. By using static semi-factual generation and dynamic human-intervened correction, RDL exploits rationales (i.e. phrases that cause the prediction), human interventions and semi-factual augmentations to decouple spurious associations and bias models towards generally applicable underlying distributions, which enables fast and accurate generalisation. Experimental results show that RDL leads to significant prediction benefits on both in-distribution and out-of-distribution tests compared to many state-of-the-art benchmarks-especially for few-shot learning scenarios. We also perform extensive ablation studies to support in-depth analyses of each component in our framework.

Semi-factual Explanations in AI

“…Other work has used similarity to a Nearest Unlike Neighbor (NUN) (Cummins and Bridge 2006) or surrogate models (like LIME) (Nugent, Doyle, and Cunningham 2009) to compute the semifactuals. More recently, Kenny & Keane (2021) advanced a generative method for computing both semi-factuals and counterfacutals in a unified framework, work that has significantly fueled interest in new uses of semi-factuals (e.g., (Artelt and Hammer 2022;Lu et al 2022;Vats et al 2022;Zhao et al 2022;Kenny and Huang 2023)).…”

Out-of-Distribution Generalization in Natural Language Processing: Past, Present, and Future

“…Data variance can be seen as a typical problem of domain generalization methods, assuming the unavailability of labeled or unlabeled data from the target domain. Previous studies have explored this approach in sentiment analysis (SA) (Kaushik et al, 2019;Ni et al, 2019;Lu et al, 2022), natural language inference (NLI) (Williams et al, 2018;Hendrycks et al, 2020), and named entity recognition (NER) (Jia and Zhang, 2020;Plank, 2021). Different domains have intrinsically different feature distributions, and instances from different domains have different predicted vocabulary distributions, which leads to the OOD generalization challenge, as shown in Figure 1.…”

“…Ideally, a model should learn rational (Jiang et al, 2021;Lu et al, 2022) features for robust generalization. Take sentiment classification for example.…”

“…Pre-trained Language Models (PLMs) (Devlin et al, 2018;Radford et al, 2018) have revolutionized natural language processing (NLP) and enabled remarkable advances in Large-scale Language Models (LLMs) (Touvron et al, 2023;Gozalo-Brizuela and Garrido-Merchan, 2023;Pichai, 2023) . Despite substantial progress in developing accurate models in several natural language understanding tasks, including sentiment analysis (Kaushik et al, 2019;Ni et al, 2019;Lu et al, 2022;Luo et al, 2022a,b), natural language inference (Williams et al, 2018), and machine reading comprehension (Kaushik and Lipton, 2018;Sugawara et al, 2020), a major challenge persists -out-of-distribution (OOD) generalization -which entails the ability of a model to 1 These authors contributed equally to this work.…”

XMD: An End-to-End Framework for Interactive Explanation-Based Debugging of NLP Models

“…Model behavior can be explained using instance (Idahl et al, 2021;Koh and Liang, 2017;Ribeiro et al, 2016) or task (Lertvittayakumjorn et al, 2020;Ribeiro et al, 2018) explanations, typically via feature importance scores. HITL feedback can be provided by modifying the explanation's feature importance scores (Kulesza et al, 2009(Kulesza et al, , 2015Zylberajch et al, 2021) or deciding the relevance of high-scoring features (Lu et al, 2022;Kulesza et al, 2010;Ribeiro et al, 2016;Teso and Kersting, 2019). The model can be updated by directly adjusting the model parameters (Kulesza et al, 2009(Kulesza et al, , 2015Smith-Renner et al, 2020), improving the training data (Koh and Liang, 2017;Ribeiro et al, 2016;Teso and Kersting, 2019), or influencing the training process (Yao et al, 2021;Cho et al, 2019;.…”