Explainable requires a quantum leap in capability and the implementation of domain specific artificial intelligence.

n Essay: Next Steps for AIA recent article by Ben Dickson published at VentureBeat.com (January 2, 2021) carries the title Leading computer scientists debate the next steps for AI in 2021.The article begins with the statement “a decade’s worth of research and application has made it clear that in its current state, deep learning is not the final solution to solving the ever-elusive challenge of creating human-level AI.”This statement is comparable to a statement by Karen Hao in the MIT Technology review, January 25, 2019, “Our study of 25 years of artificial-intelligence research suggests the era of deep learning may come to an end.”This is not unlike the conclusion attributed to Yoshua Bengio in an article by Bill Knight in Wired (October 8, 2019) (https://www.wired.com/story/ai-pioneer-algorithms-understand-why/),“Now, Bengio says deep learning needs to be fixed. He believes it won’t realize its full potential, and won’t deliver a true AI revolution, until it can go beyond pattern recognition and learn more about cause and effect. In other words, he says, deep learning needs to start asking why things happen.”Nevertheless, it is easy to find many articles that promote what may be new applications of AI ornew technologies within the field of AI. Quoting from an article by David Vergun in defense.gov January 22, 2021, “The first wave involved symbolic AI, which has explicit rules, such as if it's raining, then bring an umbrella, he said. Commercial income tax programs operate this way, using rules, logic and reasoning to reach a conclusion. The second wave involved neural nets, which Scherlis refers to as statistical AI. Neural nets attempt to replicate higher-order human thinking skills, such as problem solving. All AI relies on having good data. But although data is certainly important, the real game-changer for AI will be the third wave where symbolic is meshed with statistical to get the best of both worlds, Scherlis predicted.”This is an overly simplistic statement. It fails to identify what variety of symbolic (reasoning), how it will be implemented, what relation it would have with statistics, and more important, in what field of use will that symbolic reasoning be applied.The article by Dickson includes two important statements. First: “Cognitive scientist Gary Marcus, who cohosted the debate, reiterated some of the key shortcomings of deep learning, including excessive data requirements, low capacity for transferring knowledge to other domains, opacity, and a lack of reasoning and knowledge representation.” (emphasis mine) andSecond: “Marcus, who is an outspoken critic of deep learning–only approaches, published a paper in early 2020 in which he suggested a hybrid approach that combines learning algorithms with rules-based software.”This is reminiscent of a joke in the early ‘90s. Now that we have failed to produce anything of importance with artificial intelligence, we have now begun to study parallel artificial intelligence.While the complaint of a lack of reasoning and knowledge representation, is likely true. The conclusion that rules-based software is needed is unsupported, given the historical record of “knowledge based systems” as promoted in “The Rise of the Expert Company” (Feigenbaum, McCorduck, and Nii, 1988).

Page 2

The wired article cited above, includes the statement, “A small group of researchers is working to combine causality and machine learning.” With reference to the textbook by Judea Pearl entitled “Causality: Models, Reasoning, and Inference.” This is clearly a step in the right direction, but overlooks, the wide, technically complex, range of subjects in the 400-page textbook and how they would be applied to common everyday requirements.An alternative that seeks to avoid the weakness of “rule-based expert systems” and the unbounded complexity of causal reasoning is to focus on what may be labeled: domain specific artificial intelligence. Where it is possible to create sound logical frameworks for commonly occurring subjects and circumstances.The difference between rule-base-expert systems and domain-specific-artificial intelligence is that domain specific artificial intelligence is (or can be) composed of sound logical statements that, in effect, comprise a theory for the subject domain. And when complete, the integrity of the framework can be analyzed by humans and tested by machines to verify that they are complete and correctFor example, in the past, sequencing of factory operations based upon precedence constraints was accomplished by simple analysis of the combined (conjunctive) precedence constraints. Thislimited framework would lead to a feasible ordering of the work but fail to produce other more flexible alternative ordering. The problem is that analysis of a conjunction of precedence constraints will lead to only one solution. Strict logical analysis of conjunctive, disjunctive, and negative precedence constraints will (if feasible) lead to many alternative solutions. Moreover, analysis of conjunctive, disjunctive, and negative precedence constraints can be used to explain how the orderings were created, how the orderings compare, and how to modify the orderings during performance of work. In other words, logical analysis of conjunctive, disjunctive, and negative precedence provides both a framework for both knowledge representation and logical reasoning.This is a small, simple example of domain specific artificial intelligence that can be employed in all domains where the ordering of steps or operations must be determined in advance of action. Other simple logical frameworks can be implemented for testing, monitoring, and diagnosis of physical systems. More generally, logical frameworks can be developed for many other fields including some where the framework is understandable but the step by step analysis and inference is beyond the range of human ability.(see publications available through researchgate.net)

(PDF) An Essay - Next Steps for AI. Available from: https://www.researchgate.net/publication/349304263_An_Essay_-_Next_Steps_for_AI [accessed Apr 25 2021].

I would recommend checking out http://heatmapping.org/ , their work on LRP to explain AI is fascinating! You can even try out the interactive examples. I hope this helps. Cheers, Raoul

https://www.aclweb.org/anthology/P19-1560/

Preprint A Survey of the State of Explainable AI for Natural Language...

https://ai.googleblog.com/2020/11/the-language-interpretability-tool-lit.html

You could refer to the above links. They will surely help you in your project.

In the last month I made a literature research regarding XAI (explainable AI) and I can recommend you to XAI overview paper:

Article Peeking Inside the Black-Box: A Survey on Explainable Artifi...

Article Explainable Artificial Intelligence (XAI): Concepts, Taxonom...

Moreover, I can you recommend you the online book: Interpretable Machine Learning - A Guide for Making Black Box Models Explainable: https://christophm.github.io/interpretable-ml-book/

Natural Language Processing is not my research field, so I cannot recommend you some special XAI methods for your research. But I hope that the above listed literature helps you.