Neural language models are being increasingly used for unsupervised text classification and clustering tasks, with proposals that learn vector representations from word- to document-level. We have adapted one of the latter to discover Wikipedia articles which are relevant to selected historical topics, and also to a given question and its correct answer, by exploiting not only the knowledge captured in the writing of the articles themselves, but also in their classification in wikicategories. Our goal is to automate the generation of personalized multiple-choice quizzes, with wrong alternatives to the correct answer tailored to the level of knowledge of the target user on the selected topics. The approach is shown to provide diverse and meaningful alternatives, in a way that even the absurd ones - which are included mainly for fun-do have some interesting connections to the right answers.
展开▼
