Attentive Pooling with Learnable Norms for Text Representation