簡易檢索 / 詳目顯示

回結果列表
研究生： 梁智堯
Liang, Zhi-Yao
論文名稱： 串聯代理人強化混合式網路自動摘要
Hybrid Net based Summarization with Cascading Agents
指導教授： 高宏宇
Kao, Hung-Yu
學位類別： 碩士
Master
系所名稱： 電機資訊學院 - 資訊工程學系
Department of Computer Science and Information Engineering
論文出版年： 2019
畢業學年度： 107
語文別： 英文
論文頁數： 67
中文關鍵詞： 文本摘要 、強化學習 、自然語言生成
外文關鍵詞： Text Summarization, Reinforcement Learning, Nature Language
相關次數： 點閱：126  下載：4 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 自動摘要任務一直是自然語言處理（NLP）領域的重大挑戰。由於深度神經網絡被廣泛應用於各個領域並取得了巨大成功，摘要引起了近期研究的廣泛關注。
    處理自動摘要的方法大致可分為兩種主要方法：提取和生成方式。前者透過擷取文章中最具代表性的片段來代表摘要; 而後者透過序列生成模型來產生摘要,過程是基於理解語言的潛在含義。然而，近年來研究混合兩者方式的研究也備受推廣。
    我們採用基於強化的策略學習來混和學習兩個神經網絡。藉由設計了兩個代理人來處理關於顯著性和冗餘的兩難問題。一個代理在文檔中收到片段並提取顯著性;另一個代理設法判斷片段是否包含冗餘。首先，我們準備了兩個預先訓練的模型，包括提取和生成模型。我們的代理人對每個模型進行了特定的政策，並試圖最大化總獎勵。通過串聯機制，不僅可以通過共享觀察到的參數來全局優化模型;這些代理人也能根據各自的任務進行優化。
    最後，我們的策略直觀有效，我們混合兩種網路成功地緩解個別遭遇的困境, 在實驗中，我們也證明串聯代理可以有效解決摘要任務並且避免不必要的改寫動作，相較於傳統模型更能獲得較為精煉的語句。另外我們的模型具有相當高的可塑性,在未來也能更換不同的網路架構來處理自動摘要任務。

    Automatic summarization task has been a significant challenge in the Nature Process Language (NLP) filed. Since the deep neural networks were widely applied in various domains and achieved great success, the summarization has drawn much attention from recent research.
    The approach of dealing with automatic summarization can be roughly divided into two main paradigms: extractive and abstractive manner. The former allows capturing the most representative snippets in a document while the latter generates a summary by understanding the latent meaning in a material with a language generation model. However, in recent years, combining both paradigms to complement each other has been promoted.
    Accordingly, we adopt a reinforcement-based policy learning to bridge the non-differentiable computation between these two neural networks in a hybrid way. We design two pseudo-agents to deal with the dilemma about saliency and redundancy. One agent receives snippets within a document and extracts the saliency; the other agent manages to judge whether snippets contain the redundancy. At first, we prepared two pre-trained models, including extractive and abstractive model. Our pseudo-agents conduct a specific policy over each model and attempt to maximize the total reward. Through cascading agents, not only is the model optimized globally by sharing parameters of observations; these agents are separately learning themselves.
    Finally, our strategy is intuitive and effective. Through the hybrid network, we successfully alleviate the dilemma. In the experiments, we prove that our agents are feasible to tackle the summarization task. Also, we can keep the saliency without unnecessary rewriting behavior. With comparison with the traditional summarization model, our approach obtains more concise phrases. Besides, our model is flexible, each sub-network (extractor or abstractor) can be substituted for other robust structure in the future.

    中文摘要 I Abstract II FIGURE LISTING IX TABLE LISTING X 1. INTRODUCTION 1 1.1. Background 1 1.2. Issues 6 1.2.1. Human-Subjectivity 6 1.2.2. Local Relevance 7 1.2.3. Global Independence 9 1.3. Motivation 10 1.4. Proposed Method 11 2. RELATED WORK 13 2.1. Attention Mechanism 13 2.2. Reinforcement Learning 14 2.2.1. Markov Property 15 2.2.2. Markov Decision Process 15 2.2.3. Partially Observable Markov Decision Process 16 2.2.4. Policy Learning 18 2.2.5. Actor-Critic Framework 19 2.3. Extractive Model 21 2.3.1. Hierarchical Encoder-Decoder 22 2.3.2. Pointer Network 23 2.4. Abstractive Model 24 2.4.1. Attention-Based Sequence-to-Sequence (Seq2Seq) Model 24 2.4.2. Pointer-Generator Networks 26 2.5. Hybrid Network 27 2.5.1. FastAbs 27 3. METHOD 29 3.1. Pretrained Model 30 3.2. Agents 32 3.2.1. Extractor Agent 33 3.2.2. Abstractor Agent 35 4. EXPERIMENTS 38 4.1. Dataset 38 4.1.1. CNN/Dailymail 38 4.1.2. Dataset Description 39 4.2. Evaluation Metrics 40 4.2.1. Negative Log Likelihood (NLL) 40 4.2.2. Mean Square Error (MSE) 41 4.2.3. Precision & Recall 41 4.2.4. BLEU Score 42 4.2.5. ROUGE 43 4.2.6. F1 measure 43 4.3. Labeling 44 4.4. Evaluation 44 4.4.1. Setup 44 4.4.2. Preprocessing 45 4.4.3. Beam Search 45 4.5. Evaluation Results 46 4.5.1. Comparisons 46 4.5.1. Ablation Study 48 4.5. Performance on Extractor Agent 49 5. DISCUSSION 50 5.1. Analysis of Sequence Labeling 50 5.1.1. Reward Curve with Precision & Recall 50 5.2. Learning Curve 51 5.3. Sequential Feature Testing 54 5.4. Abstractor Agent 55 5.3.1. Policy Evaluation 56 5.3.2. Reward Analysis 58 5.5. Rerank 59 6. CONCLUSIONS 61 7. REFFERENCES 62 Appendix 65

    [1] Qingyu Zhou, Nan Yang, Furu Wei, and Ming Zhou. 2017. Selective encoding for abstractive sentence summarization. In ACL 2017, pages 1095–1104.
    [2] Alexander M. Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. In EMNLP 2015, pages 379– 389.
    [3] Ramesh Nallapati, Bowen Zhou, C´ıcero Nogueira dos Santos, C¸ aglar G ¨ulc¸ehre, and Bing Xiang. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. In CoNLL 2016, pages 280–290.
    [4] Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. SummaRuNNer: A recurrent neural network based sequence model for extractive summarization of documents. In AAAI 2017, pages 3075–3081.
    [5] Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Ranking sentences for extractive summarization with reinforcement learning. In NAACL 2018, pages 1747–1759.
    [6] Yen-Chun Chen and Mohit Bansal. 2018. Fast abstractive summarization with reinforce-selected sentence rewriting. In ACL 2018, pages 1747–1759.
    [7] Sebastian Gehrmann, Yuntian Deng, and Alexander Rush. 2018. Bottom-up abstractive summarization. In EMNLP 2018.
    [8] Qingyu Zhou, Nan Yang, Furu Wei, Shaohan Huang, Ming Zhou, and Tiejun Zhao. 2018. Neural document summarization by jointly learning to score and select sentences. In ACL 2018, pages 654–663.
    [9] Qingyu Zhou, Nan Yang, Furu Wei, Shaohan Huang, Ming Zhou, and Tiejun Zhao. 2018. Neural document summarization by jointly learning to score and select sentences. In ACL 2018, pages 654–663.
    [10] Yann N. Dauphin, Angela Fan, Michael Auli and David Grangier. 2017. Language modeling with gated convolutional networks. In ICML 2017, pages 933–941.
    [11] K. Papineni, S. Roukos, T. Ward, and W. J. Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In ACL 2002, pages 311-318.
    [12] Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In ACL Workshop 2004, pages 74–81.
    [13] Sutskever, O. Vinyals, and Q. V. Le. 2014. Sequence to sequence learning with neural networks. In NIPS 2014, pages 3104–3112.
    [14] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. 2015. Effective approaches to attentionbased neural machine translation. In EMNLP 2015, pages 1412-1421.
    [15] Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. In EMNLP 2017, pages 670-680.
    [16] O Vinyals, M Fortunato, and N Jaitly. Pointer networks. In NIPS, 2015, pages 2692–2700
    [17] Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer generator networks. In ACL 2017, pages 1073–1083.
    [18] Shi, Jiaxin; Liang, Chen; Hou, Lei, Li, Juanzi; Liu, Zhiyuan and Zhang, Hanwang. 2018. DeepChannel: Salience Estimation by Contrastive Learning for Extractive Document Summarization. In AAAI 2019.
    [19] Romain Paulus, Caiming Xiong, and Richard Socher. 2018. A deep reinforced model for abstractive summarization. In ICLR 2018, pages 1073-1083.
    [20] Haoyu Zhang, Jianjun Xu and Ji Wang. 2019.Pretraining-Based Natural Language Generation for Text Summarization. arXiv:1902.09243.
    [21] Logan Lebanoff , Kaiqiang Song and Franck Dernoncourt et al.2019. Scoring Sentence Singletons and Pairs for Abstractive Summarization. arXiv:1906.00077.
    [22] Yue Dong, Yikang Shen, Eric Crawford, Herke van Hoof, and Jackie Chi Kit Cheung.2018. Banditsum: Extractive summarization as a contextual bandit. In EMNLP 2018, pages 3739–3748.
    [23] Asli C¸ elikyilmaz, Antoine Bosselut, Xiaodong He, and Yejin Choi. 2018. Deep communicating agents for abstractive summarization. In NAACL 2018, pages 1662-1675.
    [24] Aishwarya Jadhav and Vaibhav Rajan.2018. Extractive Summarization with SWAP-NET: Sentences and Words from Alternating Pointer Networks. In ACL 2018, pages 142-155.
    [25] Ramakanth Pasunuru and Mohit Bansal. 2018. Multireward reinforced summarization with saliency and entailment. In NAACL 2018, pages 646-653.
    [26] Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. Summarunner: A recurrent neural network based sequence model for extractive summarization of documents. In AAAI 2017, pages 3075-3081.
    [27] Mihalcea, R. and P. Tarau. 2004. TextRank: Bringing order into texts. In EMNLP 2004, pages 404-441.

    下載圖示 校內：2020-08-15公開
    校外：2020-08-15公開
    QR CODE