Research On Sentence Ordering Based On Neural Network

Modeling text coherence is an essential task in natural language processing,which contributes to many downstream tasks such as extractive summarization,retrievalbased question answering and text planning.As a challenging subtask,sentence ordering aims to reconstruct a coherent paragraph from an unordered set of sentences.By learning to order sentences,the model is able to identify crucial properties that cause text coherence including logical consistency and topic transitions.Recently,inspired by the success of deep learning,neural network based models have achieved state-of-theart performance.Particularly,the sequence-to-sequence model has become a dominant framework in the sentence ordering task.In this thesis,we improve the neural sentence ordering model from two aspects of the encoder and decoder.For the encoder,we propose a paragraph encoder based on sentence-entity graphs.The sentence-entity graphs is graph-based paragraph representation involving sentence nodes and entity nodes,which can reflect entitybased coherence.Then,we introduce and improve graph recurrent networks to learn graph representations.Further,we introduce external lexical knowledge to enrich the graph structure.For the decoder,the pointer network only exploits the noisy left-side encoded context while ignoring other useful information,due to its autoregressive structure that produces the ordering sequence in a left-to-right fashion.Therefore,we propose to enhance the pointer network decoder using two pairwise ordering prediction modules,which model local coherence and relative orientations between unordered sentences respectively.On several commonly-used datasets,the proposed models significantly outperform other baselines,demonstrating the effectiveness and superiority of our method.We also evaluate our sentence ordering models on a downstream task,multidocument summarization,and the summaries reordered by our model achieve the higher coherence scores.