Social recommender systems capitalise on product reviews to generate recommendations that are both guided by experiential knowledge and are explained by user opinions centred on important product aspects. Therefore, having an effective aspect extraction algorithm is crucial. Previous work has shown that dependency relation approaches perform well in this task. However, they can also lead to erroneous extractions. This paper proposes an effective aspect extraction approach that combines strengths of both dependency relations and frequent noun approaches. Further, we demonstrate how aspect-level sentiment analysis can be used to enrich product representations and thereby positively impact recommendation effectiveness. We empirically evaluate our proposed approach with the objective to recommend products that are ‘better’ than a given query product. A computational measure of ‘better’ is used in our experiments with five real-world datasets. Results show that our proposed approach achieves significantly better results than the existing state-of-the-art dependency-based methods in recommendation tasks.