Back

SummaRuNNer: A Recurrent Neural Network Based Sequence Model for Extractive Summarization of Documents

Ramesh NallapatiFeifei ZhaiBowen Zhou

article2016AAAI1,343 citationsOutstanding Paper

Introduces SummaRuNNer, a recurrent neural network for extractive document summarization that trains directly on human-written abstractive summaries without sentence-level labels while breaking down sentence selection into interpretable factors like salience, content, and novelty.

Listen

Automated document summarization is essential for managing large volumes of textual data in information retrieval and natural language processing. While abstractive summarization generates new text, extractive summarization selects key existing sentences directly from the source text. Extractive techniques remain highly practical because they are computationally less demanding and naturally preserve grammatical and factual correctness. However, many existing deep learning approaches operate as black boxes and require costly, manually labeled sentence datasets for training.

The article introduces and evaluates SummaRuNNer, an interpretable recurrent neural network designed for extractive single-document summarization. The model processes documents sequentially to decide which sentences to include based on clear linguistic factors, and it introduces a training framework capable of learning directly from human-written summaries without requiring sentence-level labels.

The researchers designed a two-layer neural sequence classifier that models text at both the word and sentence levels. Sentence selection decisions explicitly incorporate measurable factors, including information content, document-level salience, redundancy relative to previously chosen sentences, and sentence positioning. The model was trained and evaluated on large-scale datasets, including the CNN and Daily Mail corpora containing over 280,000 news articles, and further tested on the out-of-domain DUC 2002 dataset comprising 567 documents. The authors benchmarked the system against baseline heuristics, traditional graph-based methods, and leading neural extractive and abstractive models.

The findings demonstrate three core results. First, on the Daily Mail benchmark at short summary lengths (75 bytes), SummaRuNNer achieved top-tier performance, outperforming previous deep learning models with significant improvements across standard evaluation metrics (such as a 15% relative improvement in unigram overlap over earlier leading extractive networks). Second, on the combined CNN and Daily Mail benchmark, the extractive model significantly surpassed state-of-the-art abstractive systems, establishing a clear performance advantage over generative approaches. Third, the model's transparent architecture successfully enables visual interpretability, allowing users to inspect exactly how content, salience, novelty, and position contribute to each sentence selection. Additionally, the novel training approach that learns directly from human reference summaries performed competitively, though it trailed direct extractive training by a small margin.

These results show that high-performing summarization can be achieved without the high computational complexity, hallucination risks, or large annotation expenses associated with alternative neural systems. By avoiding the need for dedicated sentence-level manual labeling, organizations can substantially reduce data preparation costs. Furthermore, the ability to decompose and visualize sentence scores directly addresses risk and compliance needs in settings where algorithmic decisions must be explainable.

Organizations deploying automated summarization should consider extractive sequence models as an efficient, low-risk alternative to generative text models. For future technical development, the source suggests combining extractive and abstractive architectures, such as using abstractive models for pre-training or building joint hybrid pipelines where extractive selections feed into text generation modules.

A key operational limitation is domain transferability: when tested on out-of-domain data like the DUC 2002 dataset, the supervised neural model performed below traditional unsupervised graph algorithms, indicating sensitivity to domain shifts. While confidence in the model's performance on news-style corpora is high, decision-makers should exercise caution and conduct domain-specific testing before deploying the system in distinct textual domains.

IBMclassificationinformation contentsrecurrent neural networksSummaRuNNerextractive summarizationdocument summarization
arXiv: 1611.04230
  • Paper: Abstractive Text Summarization using Sequence-to-sequence RNNs and Beyond, Ramesh Nallapati et al. (2016). This foundational work adapts attentional sequence-to-sequence RNNs and hierarchical document representations for summarization on the CNN/Daily Mail corpus, establishing the neural summarization framework that SummaRuNNer adapts to the extractive setting.
  • Paper: A Neural Attention Model for Abstractive Sentence Summarization, Alexander M. Rush et al. (2015). This paper pioneered the use of neural attention and encoder-decoder networks for text summarization, establishing the core sequence-modeling principles that underpin neural sentence scoring in SummaRuNNer.
  • Paper: Document Modeling with Gated Recurrent Neural Network for Sentiment Classification, Duyu Tang et al. (2015). It introduces a hierarchical gated recurrent neural architecture for composing word vectors into sentences and documents, which directly informs SummaRuNNer's two-level RNN document representation.
  • Paper: Teaching Machines to Read and Comprehend, Karl Moritz Hermann et al. (2015). This paper introduces the large-scale CNN/Daily Mail corpus and reading comprehension tasks that provided the empirical benchmark used for training and evaluating SummaRuNNer.
  • Paper: LexRank: Graph-based Lexical Centrality as Salience in Text Summarization, Günes Erkan et al. (2004). Understanding this classical graph-based lexical centrality baseline is essential for appreciating the extractive sentence-salience modeling that SummaRuNNer reformulates into an end-to-end neural sequence framework.
  • Paper: ROUGE: A Package for Automatic Evaluation of Summaries, Chin-Yew Lin (2004). This paper defines the standard ROUGE metric suite required to understand the evaluation methodology and automated reward signals used in SummaRuNNer.
  • Paper: A trainable document summarizer, J. Kupiec et al. (1995). This foundational paper establishes the statistical feature-based paradigm (salience, position, and length) for extractive sentence classification that SummaRuNNer explicitly models as interpretable neural components.
Cover for SummaRuNNer: A Recurrent Neural Network Based Sequence Model for Extractive Summarization of Documents

Abstract

We present SummaRuNNer, a Recurrent Neural Network (RNN) based sequence model for extractive summarization of documents and show that it achieves performance better than or comparable to state-of-the-art. Our model has the additional advantage of being very interpretable, since it allows visualization of its predictions broken up by abstract features such as information content, salience and novelty. Another novel contribution of our work is abstractive training of our extractive model that can train on human generated reference summaries alone, eliminating the need for sentence-level extractive labels.

Topics

classificationinformation contentsrecurrent neural networksvisualizationsentence extractionfeedforward neural networksautomatic summarization

Keywords

SummaRuNNerextractive summarizationdocument summarizationrecurrent neural network based sequence modelbidirectional RNNsentence-level RNNsequence classificationsentence representationsabstractive summarization

Table of Contents

  • Introduction
  • SummaRuNNer
  • Extractive Training
  • Abstractive Training
  • Related Work
  • Experiments and Results
  • Corpora
  • Evaluation
  • Baselines
  • SummaRuNNer Settings
  • Results on Daily Mail corpus
  • Results on CNN/Daily Mail corpus
  • Results on the Out-of-Domain DUC 2002 corpus
  • Qualitative Analysis
  • Conclusion
  • References

Knowls

  1. Knowl 1 — Hierarchical Sentence and Document Encoding in SummaRuNNer

    model/method

    SummaRuNNer models a document through a two-layer bi-directional Gated Recurrent Unit (GRU) network that operates sequentially across words and sentences.

    The lower layer operates at the word level. For sentence jj, a forward GRU and a backward GRU process word embeddings xj,tx_{j,t} to produce sequence hidden states. The sentence representation is formed by average pooling the concatenated forward and backward word hidden states.

    The upper layer operates at the sentence level. A forward GRU and a backward GRU process the pooled sentence representations across the document, yielding forward hidden state hjfh_j^f and backward hidden state hjbh_j^b at each sentence index j∈{1,…,Nd}j \in \{1, \dots, N_d\}, where NdN_d is the total sentence count. The final sentence representation hjh_j is obtained via a non-linear transformation of the concatenated bidirectional states [hjf,hjb][h_j^f, h_j^b].

    The entire document representation dd is computed by average pooling the concatenated sentence-level bidirectional hidden states followed by a non-linear projection:

    d=tanh⁡(Wd1Nd∑j=1Nd[hjf,hjb]+b)d = \tanh\left( W_d \frac{1}{N_d} \sum_{j=1}^{N_d} [h_j^f, h_j^b] + b \right)

    where WdW_d is a trainable weight matrix, bb is a bias vector, and [⋅,⋅][\cdot, \cdot] denotes vector concatenation.

  2. Knowl 2 — Interpretable Sentence Classification Function and Summary State Formulation

    equation

    In SummaRuNNer, the classification probability P(yj=1∣hj,sj,d)P(y_j = 1 \mid h_j, s_j, d) that sentence jj belongs to the summary is computed in a second sequential pass using a logistic layer that decomposes the decision into distinct linguistic and structural features:

    P(yj=1∣hj,sj,d)=σ(Wchj+hjTWsd−hjTWrtanh⁡(sj)+Wappja+Wrppjr+b)P(y_j = 1 \mid h_j, s_j, d) = \sigma\left( W_c h_j + h_j^T W_s d - h_j^T W_r \tanh(s_j) + W_{ap} p_j^a + W_{rp} p_j^r + b \right)

    where:

    • yj∈{0,1}y_j \in \{0, 1\} is a binary variable denoting summary membership for sentence jj.
    • hjh_j is the sentence representation derived from the sentence-level bi-directional GRU.
    • dd is the global document representation.
    • sjs_j is the dynamic summary representation accumulated up to sentence j−1j-1, squashed by tanh⁡\tanh to stabilize magnitude:

    sj=∑i=1j−1hiP(yi=1∣hi,si,d)s_j = \sum_{i=1}^{j-1} h_i P(y_i = 1 \mid h_i, s_i, d)

    • WchjW_c h_j models sentence content richness (WcW_c is a parameter matrix).
    • hjTWsdh_j^T W_s d captures sentence salience with respect to the overall document (WsW_s is a bilinear weight matrix).
    • −hjTWrtanh⁡(sj)-h_j^T W_r \tanh(s_j) penalizes redundancy and captures novelty relative to previously accumulated summary content (WrW_r is a bilinear weight matrix).
    • WappjaW_{ap} p_j^a captures absolute positional importance, where pjap_j^a is a learnable embedding of the sentence index jj.
    • WrppjrW_{rp} p_j^r captures relative positional importance, where pjrp_j^r is a learnable embedding of the quantized document segment ID.
    • bb is a scalar bias, and σ(z)=1/(1+e−z)\sigma(z) = 1 / (1 + e^{-z}) is the logistic sigmoid function.

  3. Knowl 3 — Greedy ROUGE Maximization for Extractive Ground-Truth Label Generation

    algorithm

    Because summarization corpora typically supply only abstractive reference summaries, SummaRuNNer uses an unsupervised greedy search to convert abstractive summaries into sentence-level binary extractive labels yj∈{0,1}y_j \in \{0, 1\}.

    Input: Document sentences D={sent1,sent2,…,sentNd}D = \{sent_1, sent_2, \dots, sent_{N_d}\}, abstractive gold summary SrefS_{ref}
    Output: Extractive subset of sentences S∗S^*
    S∗←∅S^* \leftarrow \emptyset
    scorecurr←0score_{curr} \leftarrow 0
    loop:
        best_sent←nullbest\_sent \leftarrow \text{null}
        best_score←scorecurrbest\_score \leftarrow score_{curr}
        for each sentence sent∈D∖S∗sent \in D \setminus S^*:
            cand_summary←S∗∪{sent}cand\_summary \leftarrow S^* \cup \{sent\}
            score←Rouge(cand_summary,Sref)score \leftarrow \text{Rouge}(cand\_summary, S_{ref})
            if score>best_scorescore > best\_score:
                best_score←scorebest\_score \leftarrow score
                best_sent←sentbest\_sent \leftarrow sent
        if best_sent≠nullbest\_sent \neq \text{null}:
            S∗←S∗∪{best_sent}S^* \leftarrow S^* \cup \{best\_sent\}
            scorecurr←best_scorescore_{curr} \leftarrow best\_score
        else:
            break loop
    return S∗S^*

    Sentences selected into S∗S^* receive label yj=1y_j = 1, and all others receive yj=0y_j = 0. Given training documents d∈{1,…,N}d \in \{1, \dots, N\}, the model is trained by minimizing the negative log-likelihood of these labels:

    l(W,b)=−∑d=1N∑j=1Nd(yjdlog⁡P(yjd=1∣hjd,sjd,dd)+(1−yjd)log⁡(1−P(yjd=1∣hjd,sjd,dd)))l(W, b) = -\sum_{d=1}^N \sum_{j=1}^{N_d} \left( y_j^d \log P(y_j^d = 1 \mid h_j^d, s_j^d, d_d) + (1 - y_j^d) \log(1 - P(y_j^d = 1 \mid h_j^d, s_j^d, d_d)) \right)

  4. Knowl 4 — Abstractive End-to-End Training for Extractive Summarization

    model/method

    SummaRuNNer can be trained directly from human-written reference summaries without generating synthetic extractive labels by coupling it with an auxiliary RNN decoder during training.

    The final dynamic summary state s−1=∑i=1NdhiP(yi=1∣hi,si,d)s_{-1} = \sum_{i=1}^{N_d} h_i P(y_i = 1 \mid h_i, s_i, d) computed across all document sentences serves as the context vector for the decoder GRU:

    uk=σ(Wux′xk+Wuh′hk−1+Wuc′s−1+bu′)u_k = \sigma(W'_{ux} x_k + W'_{uh} h_{k-1} + W'_{uc} s_{-1} + b'_u)

    rk=σ(Wrx′xk+Wrh′hk−1+Wrc′s−1+br′)r_k = \sigma(W'_{rx} x_k + W'_{rh} h_{k-1} + W'_{rc} s_{-1} + b'_r)

    hk′=tanh⁡(Whx′xk+Whh′(rk⊙hk−1)+Whc′s−1+bh′)h'_k = \tanh(W'_{hx} x_k + W'_{hh} (r_k \odot h_{k-1}) + W'_{hc} s_{-1} + b'_h)

    hk=(1−uk)⊙hk′+uk⊙hk−1h_k = (1 - u_k) \odot h'_k + u_k \odot h_{k-1}

    where kk indexes summary word tokens, xkx_k is the word embedding of the previous token, and primes indicate decoder parameters. Vocabulary emission probabilities Pv(wk)P_v(w_k) are produced via:

    fk=tanh⁡(Wfh′hk+Wfx′xk+Wfc′s−1+bf′)f_k = \tanh(W'_{fh} h_k + W'_{fx} x_k + W'_{fc} s_{-1} + b'_f)

    Pv(w)k=softmax(Wv′fk+bv′)P_v(w)_k = \text{softmax}(W'_v f_k + b'_v)

    The network is trained end-to-end by minimizing the negative log-likelihood of the NsN_s words in the reference summary:

    l(W,b,W′,b′)=−∑k=1Nslog⁡Pv(wk)l(W, b, W', b') = -\sum_{k=1}^{N_s} \log P_v(w_k)

    Because s−1s_{-1} is the only information channel connecting SummaRuNNer to the decoder, optimizing the generation of abstractive words forces the extractive component to learn accurate sentence extraction probabilities P(yj)P(y_j). At test time, the decoder is discarded and extractive sentence probabilities P(yj)P(y_j) are used directly.

  5. Knowl 5 — Extractive Summarization Performance on the Daily Mail Test Set

    data/table

    Performance of extractively trained SummaRuNNer and abstractively trained SummaRuNNer-abs compared against baselines on the Daily Mail test set (10,396 documents) using limited-length ROUGE recall at 75 bytes and 275 bytes:

    75 bytes Recall 275 bytes Recall
    Model Rouge-1 Rouge-2 Rouge-L Rouge-1 Rouge-2 Rouge-L
    Lead-3 21.9 7.2 11.6 40.5 14.9 32.6
    LReg (500) 18.5 6.9 10.2 - - -
    Cheng et al. (2016) 22.7 8.5 12.5 42.2 17.3 34.8*
    SummaRuNNer-abs 23.8 9.6 13.3 40.4 15.5 32.0
    SummaRuNNer 26.2 ±\pm 0.4* 10.8 ±\pm 0.3* 14.4 ±\pm 0.3* 42.0 ±\pm 0.2 16.9 ±\pm 0.4 34.1 ±\pm 0.3

    Asterisks indicate statistical significance at a 95% confidence interval with respect to the nearest model. At 75 bytes, extractively trained SummaRuNNer significantly outperforms all competing models. At 275 bytes, SummaRuNNer is statistically on par with Cheng et al. (2016) on ROUGE-1 and ROUGE-2, indicating that SummaRuNNer is especially strong at identifying the most important sentence.

  6. Knowl 6 — Full-Length Summarization Performance on the CNN/Daily Mail Test Set

    data/table

    Evaluation of extractive and abstractive models on the joint CNN/Daily Mail test set (11,480 documents) using full-length ROUGE F1 metrics:

    Model Rouge-1 F1 Rouge-2 F1 Rouge-L F1
    Lead-3 39.2 15.7 35.5
    Nallapati et al. (2016) (Abstractive) 35.4 13.3 32.6
    SummaRuNNer-abs 37.5 14.5 33.4
    SummaRuNNer 39.6 ±\pm 0.2* 16.2 ±\pm 0.2* 35.3 ±\pm 0.2

    Asterisks denote statistical significance at a 95% confidence interval over the nearest baseline. SummaRuNNer outperforms the abstractive sequence-to-sequence model of Nallapati et al. (2016) by 4.2 ROUGE-1 and 2.9 ROUGE-2 points, and achieves statistically significant improvements over the Lead-3 baseline on ROUGE-1 and ROUGE-2.

  7. Knowl 7 — Out-of-Domain Generalization on the DUC 2002 Corpus

    data/table

    Out-of-domain evaluation on the DUC 2002 test set (567 documents) using ROUGE recall capped at 75 words, evaluated with models trained on the Daily Mail corpus:

    Model Rouge-1 Rouge-2 Rouge-L
    Lead-3 43.6 21.0 40.2
    LReg 43.8 20.7 40.3
    ILP 45.4 21.3 42.8
    TGRAPH 48.1 24.3* -
    URANK 48.5* 21.5 -
    Cheng et al. (2016) 47.4 23.0 43.5
    SummaRuNNer-abs 44.8 21.0 41.2
    SummaRuNNer 46.6 ±\pm 0.8 23.1 ±\pm 0.9 43.03 ±\pm 0.8

    SummaRuNNer performs within the margin of error of Cheng et al. (2016) at a 95% confidence interval. Both neural supervised models underperform unsupervised graph-based methods (TGRAPH, URANK), indicating that supervised neural models suffer from domain shift when evaluated out-of-domain compared to graph-based unsupervised methods.

  8. Knowl 8 — Implementation Parameters and Inference Procedure for SummaRuNNer

    experimental setup

    SummaRuNNer is configured and trained with the following hyperparameter settings:

    • Embeddings & Vocabulary: 100-dimensional word2vec embeddings pretrained on the CNN/Daily Mail corpus; vocabulary truncated to the top 150,000 words.
    • Input Dimensions: Maximum of 100 sentences per document and 50 words per sentence.
    • Hidden States: GRU hidden state dimensions for both word-level and sentence-level bidirectional RNNs are set to 200.
    • Optimization: Mini-batch size of 64, optimized with Adadelta, gradient clipping for regularization, and early stopping based on validation loss.
    • Inference Strategy: At test time, sentences are ranked in descending order of predicted probabilities P(yj=1)P(y_j=1). Sentences are selected greedily until the byte or word threshold is reached for limited-length ROUGE evaluation. For full-length F1 evaluation, the top k=3k=3 highest probability sentences are chosen.

Coverage note — None was omitted; all key architectural components, equations, training algorithms (extractive and abstractive), hyperparameters, and experimental evaluations across Daily Mail, CNN/Daily Mail, and DUC 2002 are covered.

References

  1. 1.Bahdanau, D.; Cho, K.; and Bengio, Y. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473.
  2. 2.Cao, Z.; Chen, C.; Li, W.; Li, S.; Wei, F.; and Zhou, M. 2015. Tgsum: Build tweet guided multi-document summarization dataset. CoRR abs/1511.08417.
  3. 3.Cao, Z.; Li, W.; Li, S.; and Wei, F. 2016. Attsum: Joint learning of focusing and summarization with neural attention. arXiv preprint arXiv:1604.00125.
  4. 4.Carbonell, J., and Goldstein, J. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval, 335–336. ACM.
  5. 5.Cheng, J., and Lapata, M. 2016. Neural summarization by extracting sentences and words. 54th Annual Meeting of the Association for Computational Linguistics.
  6. 6.Chung, J.; Gülçehre, Ç.; Cho, K.; and Bengio, Y. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. CoRR abs/1412.3555.
  7. 7.Ganesan, K.; Zhai, C.; and Han, J. 2010. Opinosis: a graph-based approach to abstractive summarization of highly redundant opinions. In Proceedings of the 23rd international conference on computational linguistics, 340–348. Association for Computational Linguistics.
  8. 8.Hermann, K. M.; Kociský, T.; Grefenstette, E.; Espeholt, L.; Kay, W.; Suleyman, M.; and Blunsom, P. 2015. Teaching machines to read and comprehend. CoRR abs/1506.03340.
  9. 9.Kageback, M.; Mogren, O.; Tahmasebi, N.; and Dubhashi, D. 2014. Extractive summarization using continuous vector space models. 31–39.
  10. 10.McDonald, R. 2007. A study of global inference algorithms in multi-document summarization. 557–564.
  11. 11.Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G. S.; and Dean, J. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, 3111–3119.
  12. 12.Nallapati, R.; Zhou, B.; dos Santos, C.; Gulcehre, C.; and Xiang, B. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. The SIGNLL Conference on Computational Natural Language Learning.
  13. 13.Nallapati, R.; Zhou, B.; and Xiang, B. 2016. Sequence-to-sequence rnns for text summarization. International Conference on Learning Representations, Workshop track.
  14. 14.Parveen, D.; Ramsl, H.-M.; and Strube, M. 2015. Topical coherence for graph-based extractive summarization. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 19491954.
  15. 15.Radev, D., and Erkan, G. 2004. Lexrank: Graph-based lexical centrality as salience in text summarization. Journal of Artificial Intelligence Research 457–479.
  16. 16.Rush, A. M.; Chopra, S.; and Weston, J. 2015. A neural attention model for abstractive sentence summarization. arXiv preprint arXiv:1509.00685.
  17. 17.Shen, D.; Sun, J.-T.; Li, H.; Yang, Q.; and Chen, Z. 2007. Document summarization using conditional random fields. In Proceedings of IJCAI.
  18. 18.Socher, R.; Huang, E. H.; Pennin, J.; Manning, C. D.; and Ng, A. Y. 2011. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. 801–809.
  19. 19.Svore, K. M.; Vanderwende, L.; and Burges, C. J. 2007. Enhancing single-document summarization by combining ranknet and third-party sources. In Proceedings of the Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, 448–457.
  20. 20.Wan, X. 2010. Towards a unified approach to simultaneous single-document and multidocument summarizations. In In Proceedings of the 23rd COLING, 11371145.
  21. 21.Woodsend, K., and Lapata, M. 2010. Automatic generation of story highlights. In In Proceedings of the 48th ACL, 565574.
  22. 22.Yin, W., and Pei, Y. 2015. Optimizing sentence modeling and selection for document summarization. In Proceedings of the 24th International Conference on Artificial Intelligence, 1383–1389. AAAI Press.
  23. 23.Zeiler, M. D. 2012. ADADELTA: an adaptive learning rate method. CoRR abs/1212.5701.

Citation

MLA
Nallapati, R., et al. “SummaRuNNer: A Recurrent Neural Network Based Sequence Model for Extractive Summarization of Documents”. arXiv, 2016, http://arxiv.org/abs/1611.04230v1.
APA
Nallapati, R., Zhai, F., & Zhou, B. (2016). SummaRuNNer: A Recurrent Neural Network based Sequence Model for Extractive Summarization of Documents. arXiv. http://arxiv.org/abs/1611.04230v1
Chicago
Nallapati, R., F. Zhai, and B. Zhou. 2016. “SummaRuNNer: A Recurrent Neural Network Based Sequence Model for Extractive Summarization of Documents”. arXiv. http://arxiv.org/abs/1611.04230v1.
Harvard
Nallapati, R., Zhai, F. and Zhou, B. (2016) “SummaRuNNer: A Recurrent Neural Network based Sequence Model for Extractive Summarization of Documents”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1611.04230v1.
Vancouver
1. Nallapati R, Zhai F, Zhou B (2016) SummaRuNNer: A Recurrent Neural Network based Sequence Model for Extractive Summarization of Documents. arXiv

BibTeX

@article{nallapati2016summarunner,
  title = {SummaRuNNer: A Recurrent Neural Network based Sequence Model for Extractive Summarization of Documents},
  author = {Nallapati, Ramesh and Zhai, Feifei and Zhou, Bowen},
  year = {2016},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1611.04230v1},
  eprint = {1611.04230}
}
Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF