@inproceedings{ren2017leveraging, author = {Ren, Pengjie and Chen, Zhumin and Ren, Zhaochun and Wei, Furu and Ma, Jun and Rijke, Maarten de}, title = {Leveraging Contextual Sentence Relations for Extractive Summarization Using a Neural Attention Model}, booktitle = {SIGIR '17, Shinjuku, Tokyo, Japan, August 07-11, 2017}, year = {2017}, month = {August}, abstract = {As a framework for extractive summarization, sentence regression has achieved state-of-the-art performance in several widely-used practical systems. The most challenging task within the sentence regression framework is to identify discriminative features to encode a sentence into a feature vector. So far, sentence regression approaches have neglected to use features that capture contextual relations among sentences. We propose a neural network model, Contextual Relation-based Summarization (CRSum), to take advantage of contextual relations among sentences so as to improve the performance of sentence regression. Specifically, we first use sentence relations with a wordlevel attentive pooling convolutional neural network to construct sentence representations. Then, we use contextual relations with a sentence-level attentive pooling recurrent neural network to construct context representations. Finally, CRSum automatically learns useful contextual features by jointly learning representations of sentences and similarity scores between a sentence and sentences in its context. Using a two-level attention mechanism, CRSum is able to pay attention to important content, i.e., words and sentences, in the surrounding context of a given sentence. We carry out extensive experiments on six benchmark We propose a neural network model, Contextual Relation-based Summarization (CRSum), to take advantage of contextual relations among sentences so as to improve the performance of sentence regression. Specifically, we first use sentence relations with a wordlevel attentive pooling convolutional neural network to construct sentence representations. Then, we use contextual relations with a sentence-level attentive pooling recurrent neural network to construct context representations. Finally, CRSum automatically learns useful contextual features by jointly learning representations of sentences and similarity scores between a sentence and sentences in its context. Using a two-level attention mechanism, CRSum is able to pay attention to important content, i.e., words and sentences, in the surrounding context of a given sentence. We carry out extensive experiments on six benchmark datasets. CRSum alone can achieve comparable performance with state-of-the-art approaches; when combined with a few basic surface features, it signi€cantly outperforms the state-of-the-art in terms of multiple ROUGE metrics.}, publisher = {ACM}, url = {http://approjects.co.za/?big=en-us/research/publication/leveraging-contextual-sentence-relations-extractive-summarization-using-neural-attention-model/}, edition = {SIGIR ’17, Shinjuku, Tokyo, Japan, August 07-11, 2017}, }