Evaluating Explanation Methods for Neural Machine Translation

• DOI: 10.18653/v1/2020.acl-main.35
• Publication Date: 2020-07-29T14:14:43Z
• Authors: Jierui Li; Lemao Liu; Huayang Li; Guanlin Li; Guoping Huang; Shuming Shi
• Keywords: Black box

Recently many efforts have been devoted to interpreting the black-box NMT models, but little progress has made on metrics evaluate explanation methods. Word Alignment Error Rate can be used as such a metric that matches human understanding, however, it not measure methods those target words are aligned any source word. This paper thereby makes an initial attempt from alternative viewpoint. To this end, proposes principled based fidelity in regard predictive behavior of model. As exact computation for is intractable, we employ efficient approach its approximation. On six standard translation tasks, quantitatively several terms proposed and reveal some valuable findings these our experiments.