When More Than One Answer is Correct: Lessons for AI in Education

Recent advances in NLP have opened new opportunities for education. Yet, as we integrate NLP into language learning, we face a fundamental challenge: evaluation of learner input. How do we assess answers fairly when multiple responses may be equally valid? And how do we ensure that our metrics capture human judgments of fluency and adequacy? In this talk, I will present my research on assessing learner answers in computer-aided language learning, carried out within the Revita platform. I will focus on the phenomenon of alternative-correct answers—cases where learners provide responses that are grammatically and semantically correct but not anticipated by the system. Addressing this challenge requires bridging NLP techniques with linguistic theory and pedagogy. I will outline approaches based on grammatical error detection/correction, and probing of Transformer models, highlighting their strengths and limitations. Beyond methodology, I will reflect on broader lessons for educational NLP: the importance of building learner corpora, developing evaluation frameworks that go beyond binary correctness, and designing systems that adapt to learners rather than restrict them. I will also connect these ideas to ongoing efforts in benchmarking GEC evaluation, arguing that fair evaluation is crucial not only for educational applications but also for advancing NLP as a field.

Featured Talk 1