Keep the padding out of the summary

Texts have different lengths, while many downstream models expect a fixed number of inputs. Before building a complicated sequence model, we can ask whether a simple summary of the token vectors already captures enough for our task.