A Survey of the State of Explainable AI for Natural Language Processing

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 447–459 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 447 A Survey of the State of Explainable AI for Natural Language Processing Marina Danilevsky, Kun Qian, Ranit Aharonov, Yannis Katsis, Ban Kawas, Prithviraj Sen IBM Research – Almadem mdanile@us.ibm.com, {qian.kun,Ranit.Aharonov2}@ibm.com yannis.katsis@ibm.com, {bkawas,senp}@us.ibm.com Abstract Recent years have seen important advances in the quality of state-of-the-art models, but this has come at the expense of models becom- ing less interpretable. This survey presents an overview of the current state of Explainable AI (XAI), considered within the domain of Natural Language Processing (NLP). We dis- cuss the main categorization of explanations, as well as the various ways explanations can be arrived at and visualized. We detail the oper- ations and explainability techniques currently available for generating explanations for NLP model predictions, to serve as a resource for model developers in the community. Finally, we point out the current gaps and encourage directions for future work in this important re- search area. 1 Introduction Traditionally, Natural Language Processing (NLP) systems have been mostly based on techniques that are inherently explainable. Examples of such ap- proaches, often referred to as white box techniques, include rules, decision trees, hidden Markov mod- els, logistic regressions, and others. Recent years, though, have brought the advent and popularity of black box techniques, such as deep learning mod- els and the use of language embeddings as features. While these methods in many cases substantially advance model quality, they come at the expense of models becoming less interpretable. This ob- fuscation of the process by which a model arrives at its results can be problematic, as it may erode trust in the many AI systems humans interact with daily (e.g., chatbots, recommendation systems, in- formation retrieval algorithms, and many others). In the broader AI community, this growing under- standing of the importance of explainability has cre- ated an emerging field called Explainable AI (XAI). However, just as tasks in different fields are more amenable to particular approaches, explainability must also be considered within the context of each discipline. We therefore focus this survey on XAI works in the domain of NLP, as represented in the main NLP conferences in the last seven years. This is, to the best of our knowledge, the first XAI sur- vey focusing on the NLP domain. As will become clear in this survey, explainabil- ity is in itself a term that requires an explanation. While explainability may generally serve many purposes (see, e.g., Lertvittayakumjorn and Toni, 2019), our focus is on explainability from the per- spective of an end user whose goal is to understand how a model arrives at its result, also referred to as the outcome explanation problem (Guidotti et al., 2018). In this regard, explanations can help users of NLP-based AI systems build trust in these sys- tems’ predictions. Additionally, understanding the model’s operation may …

Content truncated to 500 words