Trusting Language Models in Education

The innovation that Deep Learning has brought in the era of Big Data is considered a breakthrough, since those models gave practitioners the ability to solve a wide collection of difficult problems that Classical Machine Learning approaches couldn't perform well [1, 2, 3, 4, 5]. For example, we have seen great improvements in the medical area using computer vision [6, 7], and also in Natural Language Processing (NLP) [8, 9], just to cite a few examples. This last area is going to be the focus of this paper. Specifically, at CK-12, we have a Question & Answering (Q&A) that starting with an input query, goes through several stages of processing. After the final stage we arrive at a set of candidate paragraphs that are likely to contain an answer to the query. This final stage is a softmax that ranks the paragraphs according to how likely they are to contain the correct answer.