Learning to Explain: Generating Stable Explanations Fast

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5340–5355 August 1–6, 2021. ©2021 Association for Computational Linguistics 5340 Learning to Explain: Generating Stable Explanations Fast Xuelin Situ† Ingrid Zukerman† Cecile Paris‡ Sameen Maruf† Gholamreza Haffari† †Dept of Data Science and AI, Faculty of IT, Monash University, Australia ‡CSIRO Data61, Australia †{firstname.lastname}@monash.edu ‡{firstname.lastname}@data61.csiro.au Abstract The importance of explaining the outcome of a machine learning model, especially a black- box model, is widely acknowledged. Recent approaches explain an outcome by identifying the contributions of input features to this out- come. In environments involving large black- box models or complex inputs, this leads to computationally demanding algorithms. Fur- ther, these algorithms often suffer from low stability, with explanations varying signifi- cantly across similar examples. In this pa- per, we propose a Learning to Explain (L2E) approach that learns the behaviour of an un- derlying explanation algorithm simultaneously from all training examples. Once the explana- tion algorithm is distilled into an explainer net- work, it can be used to explain new instances. Our experiments on three classification tasks, which compare our approach to six explana- tion algorithms, show that L2E is between 5 and 7.5 × 104 times faster than these algo- rithms, while generating more stable explana- tions, and having comparable faithfulness to the black-box model. 1 Introduction Explaining the mechanisms and reasoning behind the outcome of complex machine learning models, such as deep neural networks (DNNs), is crucial. Such explanations can shed light on the potential flaws and biases within these powerful and widely applicable models, e.g., in medical diagnosis (Caru- ana et al., 2015) and judicial systems (Rich, 2016). Existing explainability methods mostly produce explanations, or rationales (DeYoung et al., 2020), which identify the attributions of features in an in- put example, e.g., are they contributing positively or negatively to the prediction of an outcome. For text classifiers, this means identifying words or phrases in an input document that account for a Novell’s Microsoft attack completes Linux conversion: Novell Inc. has completed its conversion to Linux by launching an attack on Microsoft Corp., claiming that the company has stifled software innovation and that the market will abandon Microsoft Windows at some point in the future. ˆyxxx = 99% Sci/Tech; ˆyxxx∖A = 14%; ˆyxxx∖L2E = 0.7% Microsoft expands Windows update Release: Microsoft Corp. is starting to ramp up distribution of its massive security update for the Windows XP operating system, but analysts say they still expect the company to move at a relatively slow pace to avoid widespread glitches. ˆyxxx = 98% Sci/Tech; ˆyxxx∖A = 66%; ˆyxxx∖L2E = 0.4% Figure 1: Two similar examples from the News dataset. The most important words (top 30%) found by our method L2E are yellow-highlighted, and those from a baseline A are underlined. L2E considers words like ‘Microsoft’ and ‘Win- dows’ important in both examples. ˆyxxx is the model’s pre- diction, and ˆyxxx∖· is the model’s prediction after removing important words …

Content truncated to 500 words