近來序列生成對抗網路火紅，在文字生成與對話生成領域都有好成果。因此，我們希望透過序列生成對抗網路建立問答系統，教導機器人如何閱讀相關文章，產生適當且擬人的答案。同時，智慧客服需求日漸增加，市面上的系統多以模板導向為主，過度倚賴訂立的規則和資料，易有無法回應用戶的窘境，且多侷限於特定的任務。因此我們嘗試使用序列生成對抗網路建立智慧客服問答系統。
在本論文中共有三個目標。第一部份為生成器與辨別器之事先訓練。生成器屬於序列對序列模型，使用長短期記憶網路進行訓練，有別於其他研究多採取固定字典之方法，我們通過動態字典降低解碼器選擇錯誤字詞的機率，藉此產生更好的結果。辨別器使用卷積式網路進行文字分類訓練，能對完整或部分序列進行評分。第二部份開始序列生成對抗網路，由於生成的答案和實際答案可能長度差異極大，因此在計算回饋時將長度納入考量，實驗結果證明的確能提升效能。最終我們替企業建立智慧客服問答系統，用以回答產品的相關問題。因此我們以企業網站上的產品介紹、使用手冊等資訊做為知識背景，並使用企業提供的FAQ進行訓練。

Recently, Sequence Generative Adversarial Nets is very popular. It has achieved good results in both text generation and dialog generation. Therefore, we planned to build a question answering system via SeqGAN to teach machine reading question-related documents and generate answers like human responses. The demands for smart customer service is increasing day by day. Most of the systems built by template-based method failed to response users easily whenever the question asked by user does not follow the established rule or does not exist in the database, hence they are often limited on specific tasks. To resolve the weaknesses of the template-based customer service, we try to build a customer service with question answering system based on SeqGAN.
This thesis involves three parts. The first part is the pre-training of the generator and the discriminator. The generator is a seq2seq model with GRU. Besides, we use dynamic dictionary to decrease the probability that the decoder generates wrong words. The discriminator, which is a CNN model to classify the QA pair generated by human or machine, could give reward for both fully and partially QA pairs. The second part is SeqGAN. Different from general generative adversarial training, we consider the length difference between actual answer and generated answer when calculating the reward. Experimental results proved that the proposed system improved the performance successfully. In the last part we try to build a customer service system. We crawl the information such as product information and user manuals on the company website and use FAQ provided by the company to train our model.

[1] UNGER, Christina, et al. Template-based question answering over RDF data. In: Proceedings of the 21st international conference on World Wide Web. ACM, 2012. p. 639-648.
[2] FENG, Minwei, et al. Applying deep learning to answer selection: A study and an open task. In: Automatic Speech Recognition and Understanding (ASRU), 2015 IEEE Workshop on. IEEE, 2015. p. 813-820.
[3] TAN, Ming, et al. LSTM-based deep learning models for non-factoid answer selection. arXiv preprint arXiv:1511.04108, 2015.
[4] MIKOLOV, Tomas, et al. Distributed representations of words and phrases and their compositionality. In: Advances in neural information processing systems. 2013. p. 3111-3119.
[5] SUTSKEVER, Ilya; VINYALS, Oriol; LE, Quoc V. Sequence to sequence learning with neural networks. In: Advances in neural information processing systems. 2014. p. 3104-3112.
[6] YIN, Jun, et al. Neural generative question answering. arXiv preprint arXiv:1512.01337, 2015.
[7] MITRA, Rajarshee. An Abstractive approach to Question Answering. arXiv preprint arXiv:1711.06238, 2017.
[8] YU, Lantao, et al. SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient. In: AAAI. 2017. p. 2852-2858.
[9] LI, Jiwei, et al. Adversarial learning for neural dialogue generation. arXiv preprint arXiv:1701.06547, 2017.
[10] GUYEN, Tri, et al. MS MARCO: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268, 2016.
[11] MNIH, Volodymyr, et al. Asynchronous methods for deep reinforcement learning. In: International Conference on Machine Learning. 2016. p. 1928-1937.
[12] GOODFELLOW, Ian, et al. Generative adversarial nets. In: Advances in neural information processing systems. 2014. p. 2672-2680.
[13] LEDIG, Christian, et al. Photo-realistic single image super-resolution using a generative adversarial network. arXiv preprint, 2016.
[14] SHRIVASTAVA, Ashish, et al. Learning from simulated and unsupervised images through adversarial training. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017. p. 6.
[15] ZHU, Jun-Yan, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks. arXiv preprint arXiv:1703.10593, 2017.
[16] YU, Lantao, et al. SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient. In: AAAI. 2017. p. 2852-2858.
[17] LI, Jiwei, et al. Adversarial learning for neural dialogue generation. arXiv preprint arXiv:1701.06547, 2017.