SHAP-Based Explanation Methods: A Review for NLP Interpretability

Proceedings of the 29th International Conference on Computational Linguistics, pages 4593–4603 October 12–17, 2022. 4593 SHAP-Based Explanation Methods: A Review for NLP Interpretability Edoardo Mosca TU Munich, Department of Informatics, Germany edoardo.mosca@tum.de Ferenc Szigeti TU Munich, Department of Informatics, Germany ferenc.szigeti@tum.de Stella Tragianni TU Munich, Department of Informatics, Germany stella.tragianni@tum.de Daniel Gallagher University College Dublin, Department of Informatics, Ireland daniel.gallagher1@ucdconnect.ie Georg Groh TU Munich, Department of Informatics, Germany grohg@in.tum.de Abstract Model explanations are crucial for the transpar- ent, safe, and trustworthy deployment of ma- chine learning models. The SHapley Additive exPlanations (SHAP) framework is considered by many to be a gold standard for local explana- tions thanks to its solid theoretical background and general applicability. In the years following its publication, several variants appeared in the literature—presenting adaptations in the core assumptions and target applications. In this work, we review all relevant SHAP-based inter- pretability approaches available to date and pro- vide instructive examples as well as recommen- dations regarding their applicability to NLP use cases. 1 Introduction Several methods have been proposed to address the issue of opacity in modern machine learning models. Most notoriously, explanations are funda- mental for Deep Neural Networks (DNNs) (Devlin et al., 2019; Madsen et al., 2021; Mosca et al., 2021) as these automatically learn millions of pa- rameters and behave like black-boxes. Lundberg and Lee (2017) proposes SHapley Additive exPla- nations (SHAP), a unified local-interpretability framework with a rigorous theoretical foundation on the game-theoretic concept of Shapley values (Shapley, 1953). SHAP is nowadays considered a core contri- bution to the field of eXplainable Artificial Intel- ligence (XAI). Following its publication, a vari- ety of explainability approaches based on SHAP’s methodology has populated the literature and this Shapley Values (1953) Different Inputs SHAP Framework (2017) Improved Efficiency Different Assumptions Different Models Different Scope SAGE L-Shapley Surrogate SHAP SealSHAP ConceptSHAP ASV Shapr Shapley Flow TreeSHAP SubgraphX DASP HEDGE TimeSHAP h-SHAP Figure 1: This work identifies five research directions pursued by Shapley- and SHAP-based approaches in XAI. Each direction, together with a few notable meth- ods as examples, has been indicated by a different color. trend continues to grow. Some present a new ver- sion of SHAP tailored to a certain type of input data—e.g. graphs (Yuan et al., 2021) and text (Chen et al., 2020)—or to specific models such as random forests (Lundberg et al., 2018). Others, instead, modify SHAP’s underlying assumptions— e.g. features independence—to increase the origi- nal framework’s flexibility for cases in which they are too strict or overly simplistic (Frye et al., 2019). In this work, we (1) identify five broad research directions inspired by SHAP, (2) review available SHAP-based (or Shapley-value-based) approaches as members of such categories, and (3) investigate their applicability in the domain of Natural Lan- guage Processing (NLP). Our work reviews 41 methods with a particu- lar focus on their core assumptions, input require- 4594 ments, explanation form, and available implemen- tations. Furthermore, we provide NLP researchers with use-case-based recommendations and instruc- tive examples. 2 Background For the …

Content truncated to 500 words