Various methods for relation extraction and classification have been proposed and benchmarked onstandard academic datasets. In real-life data however, the number of potential relations is enormousand the heuristics and count-based methods often used to separate actual relations from meaninglessco-occurrences fail to detect weak signals of importance. In this paper, we investigate the use of acomputationally-light binary detection model to identify whether a couple of entities in a sentencebears a relation, as a preliminary step prior to relation classification. Our model is based on the shortestdependency path between two entities analyzed with a LSTM recurrent network and combined withinformation on the entities types. On the binary relation detection task, we achieve results betterthan a state-of-the-art relation classification model adapted to detection with increased robustnessto relations unseen in training. We finally show that binary detection as a pre-processing step toclassification of relations is effective in significantly improving the latter.