Verbalization of knowledge base (KB) triples about an entity allows users to absorb information from KB more easily. The drawback of most of the previous work is that they cannot generalize to unseen frames. We propose two variants of sequence-to-sequence neural networks architectures with pointer network that are able verbalize unseen frames. The results of automatic evaluation and human evaluation both indicate a local pointer network improves Meteor and slot error rates over the baseline model. Different from previous work, we construct the corpus without human annotation. We also find that current relation extraction systems are not yet able to construct high quality corpora.