Product reviews which are increasingly commonplace on the web typically contain a textual component and a numerical rating. The textual component can be viewed as a collection of arguments for and against the product. Whilst the reviewer may not have provided the attacks between these arguments they typically provide an indication of which set of arguments they view as being more acceptable/winning via the numerical rating (i.e. a positive rating indicates that the positive arguments are accepted and vice versa). Our framework builds upon this intuition and we propose a two step process for identifying a probability distribution over the set of possible argument graphs that the reviewer may have had in mind. The first is the identification step in which for a given review, we identify a distribution by analysing the relationship between the rating and polarity of arguments in the review via the constellations approach to probabilistic argumentation. The second step is the refinement step in which we harness ratings from multiple reviews and use this to refine our probability distribution thus enabling us to learn from the data. We illustrate the applicability of our approach by testing it with real data.