1: \begin{abstract}

2:   Machine learning approached through supervised learning requires expensive

3:   annotation of data. This motivates weakly supervised learning, where data are

4:   annotated with incomplete yet discriminative information. In this paper, we

5:   focus on partial labelling, an instance of weak supervision where, from a given

6:   input, we are given a set of potential targets. We review a disambiguation

7:   principle to recover full supervision from weak supervision, and propose an

8:   empirical disambiguation algorithm. We prove exponential convergence rates of

9:   our algorithm under classical learnability assumptions, and we illustrate the

10:   usefulness of our method on practical examples.

11: \end{abstract}

12: