Multi-concept Alignment and Evaluation

In this paper we discuss a book annotation translation application scenario that requires multi-concept alignment - where one set of concepts is aligned to another set. Using books annotated by concepts from two vocabularies which are to be aligned, we explore two statistically-grounded measures (Jaccard and LSA) to build conversion rules which aggregate similar concepts. Different ways of learning and deploying the multi-concept alignment are evaluated, which enables us to assess the usefulness of the approach for this scenario. This usefulness is low at the moment, but the experiment has given us the opportunity to learn some important lessons.