Significant obstacles must be overcome if machine learning techniques are to be applied in the legal domain. Our experience with the Split-Up project has led us to conclude that for machine learning to be applied usefully in legal domains, (i) the domain being modelled must be bounded and (i.i) the domain requires an abundance of commonplace cases. This research has lead us to develop strategies for using machine learning to build legal knowledge based systems. We discuss these strategies in respect to the Split-Up project. Split-Up uses machine learning to model how an Australian Family Court judge distributes marital property following divorce. In law, an explanation for a decision reached is often more important than the decision. We advocate the use of Toulmin’s theory of argumentation to provide explanations to support the outcomes predicted by our knowledge discovery system Split-Up.