Using Conditional Sentence Representation in Pointer Networks for Sentence Ordering

Sentence ordering aims to arrange sentences in a coherent manner and hence has important applications in Natural Language Generation. Recently, several approaches have used Pointer Networks for this task. Such networks arrange a list of sentences based on fixed sentence representations, where these representations are independent of the sentence's position in the text and its relation to the previously selected sentences. In this work, we propose a conditional sentence representation, which incorporates the information of the previously selected sentences into the candidate sentence representations. By using such information, the Pointer Network is able to better capture dependencies among sentences. Experiments indicate that our proposed model achieves state-of-the-art performance on most sentence ordering benchmarks and achieves a significant improvement over state-of-the-art performance on short stories datasets.