Explainable AI for English Grammar Teaching: Making Neural Decisions Transparent to Learners
Abstract
There are tools for grammar checking, grammatical error correction (GEC) and conversational tutors that are now powered by neural networks and that give verdicts without reasons. An unstated correction is pedagogically thin for learners, as they have to notice, understand and internalise rules. The paper is an integrative review based on secondary data comprised of 40 peer-reviewed and archival papers published from 1988 to 2022 covering explainable AI (XAI), neural-language-model interpretability, GEC and second-language-acquisition (SLA) studies dedicated to feedback research. These are guided by three research questions: What are the different types of explanation techniques available and how do they fit into the context of teaching grammar? What should explanations be doing, as determined by the pedagogy? What challenges exist that make the deployment of such techniques difficult? The review reveals that (a) explanation methods vary significantly in faithfulness and learner intelligibility, and explanation displayed in a learner does not necessarily constitute evidence of noticing and explicit feedback in SLA; (b) in the corpus of studies reviewed, there is very little direct experimental evidence of XAI in the context of grammar learning, and therefore, conclusions have to be largely based on neighboring evidence from AE and IT in the context of writing; and (c) direct experimental evidence of XAI in the context of grammar learning is relatively scarce in the corpus reviewed, and therefore, conclusions have to be mostly based on neighboring evidence from AE and IT in the context of writing, with (d) being the conclusion that attention-weight displays are an unreliable basis for learner facing explanations, and that unfaithful explanations and cognitive overload and over-reliance are the main risks. The paper suggests the four layer Transparent Grammar Explanation Framework (TGEF) and the Research agenda.