臺灣博碩士論文加值系統

深度學習中的序列模型經常被用來處理自然語言任務﹙Natural Language Processing, NLP﹚，而人類的語言做為模型的訓練數據，經常攜帶著文化上的偏見和歧視――例如種族歧視和性別刻板印象，這些偏見隨著數據被編碼進訓練模型中之後，容易形成機會不均的道德問題。
本研究以遞歸神經網路﹙Recurrent Neural Network, RNN﹚所架構的序列到序列﹙Sequence-to-Sequence, Seq2Seq﹚模型作為聊天機器人的基礎架構，並選取「職業種類」做為去偏見化目的，選取「性別」為去偏見化要素，利用本研究所設計的損失函數，訓練出一個對職業種類較少性別偏見的聊天機器人。
本研究設計了一問答中立性指標，用來檢測聊天機器人對不同職業的性別偏好。同時本研究亦關注除偏見方法的泛化問題，並藉此回到人類實際社會的層面進一步來探討文化中的族群歧視。
人工智慧的偏見問題的影響廣泛且深遠，其公平性問題對人類社會的影響卻容易被低估。在國外，此領域的研究亦剛起步，本研究做為臺灣探索促進公平性的技術方法的研究先例，試圖量化人類的歧視、偏見和不平等，在技術層次上加以去除和改善。

The sequential model of deep learning is commonly used to deal with Natural Language Processing (NLP) tasks. And “language” as a source of training data with human bias, e.g. racial and gender discrimination, would encode the bias into deep learning model, resulting in ethical issues about fairness.
A Sequence-to-Sequence(Seq2Seq) model with Recurrent Neural Network (RNN) is constructed for chatbots, and “occupation” and “gender” are selected as the purpose and factor of debias, respectively. A neutral index of chatbot is designed in this research for quantifying the preference of gender in the responses of occupation-related questions. As a result, we empirically demonstrate that the neutral index of debiased model is less than that of original one.
In this research, a designed term of bias loss is added into the training process for reducing human bias in deep learning model. Furthermore, generalization capability of the debiased model is concerned in the form of a “neutral words” selection problem, which can be referred back to the reality of human culture.
The fairness issues in artificial intelligence and machine learning has profound impact on this world, but it’s often underestimated. It just started to be noticed oversea in recent years. As a pioneer of Taiwan in this area, this study is an urge on the technological improvement of machine fairness issue.

一、緒論......1
1.1 研究動機......1
1.1.1人工智慧的崛起......1
1.1.2 機器決策的中立性問題......2
1.1.3自然語言的文化偏見......4
1.1.4 研究動機總結......5
1.2 研究背景......5
1.2.1去偏見化主題選擇――職業與性別......5
1.2.2 當前文獻......7
1.3 研究問題......9
1.4 研究重要性......10
二、文獻探討......11
2.1自然語言處理相關的機器學習與人類文化偏見......11
2.1.1 當前文獻架構總覽......11
2.1.2自然語言方向的當前文獻......13
2.2 詞嵌入（Word Embedding）......15
2.2.1詞向量（Word Vector）......15
2.2.2詞嵌入（Word Embedding）......16
2.3 序列生成相關的類神經網路模型......19
2.3.1序列資料的建模......19
2.3.2循環神經網路（Recurrent Neural Networks，RNN）......20
2.3.3長短期記憶（Long short-term memory，LSTM）......21
2.3.4 序列到序列條件生成模型（Sequence to Sequence Conditioned Generation Framework，Seq2Seq）......23
三、研究方法......26
3.1 研究假設......26
3.1.1針對第一個研究問題的研究假設......26
3.1.2針對第二個研究問題的研究假設......27
3.2 研究情境與模型架構......28
3.2.1研究情境......28
3.2.2模型架構......29
3.3 實驗設計與研究流程......30
3.4 資料集與資料預處理......31
3.4.1 聊天問答語料......31
3.4.1.1 資料擷取......31
3.4.1.2 資料清理......33
3.4.2 族群字與中性字......35
3.4.3 預訓練詞嵌入......36
3.5 模型實作細節......37
3.5.1 開發環境與使用套件......37
3.5.2 模型參數......38
3.5.3 訓練參數......39
3.6 研究工具......40
3.6.1 成效測量及評估方式......40
3.6.2詞嵌入類比分數（Word Embedding Analogy Score）......41
3.6.3詞嵌入偏見程度（Word Embedding Bias）......41
3.6.4聊天機器人回覆偏見程度（Chatbot Response Bias）......42
四、實驗結果與討論......44
4.1 實驗一、以職業為中性字的職業去偏見化結果......44
4.1.1 詞嵌入類比分數（Word Embedding Analogy Score）......44
4.1.2 詞嵌入偏見程度......44
4.1.3 聊天機器人回覆偏見程度......45
4.2 實驗二、以褒/貶義詞為中性字的職業去偏見化結果......46
4.2.1 詞嵌入類比分數（Word Embedding Analogy Score）......46
4.2.2 詞嵌入偏見程度......47
4.2.3 聊天機器人回覆偏見程度......48
4.3 研究問題與討論......49
五、結論與未來建議......51
5.1 結論......51
5.2 研究限制......51
5.3 未來建議......52
參考文獻......53

[1] Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, Adam Kalai, “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings.” In Conference on Neural Information Processing Systems (NIPS), 2016.
[2] Aylin Caliskan, Joanna J. Bryson, Arvind Narayanan, “Semantics derived automatically from language corpora contain human-like biases.” In Journal of Science, 356(6334):183-186, Apr 2017.
[3] Karen Ross, Cynthia Carter. “Women and news: A long and winding road.” In Journal of Media, Culture & Society, 33(8):1148–1165, 2011.
[4] Matthew Kay, Cynthia Matuszek, Sean A. Munson. “Unequal representation and gender stereotypes in image search results for occupations.” Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, pp. 3819-3828, 2015.
[5] Claudia Wagner, David Garcia, Mohsen Jadidi, Markus Strohmaier. “It’s a man’s wikipedia? assessing gender inequality in an online encyclopedia.” In Ninth International AAAI Conference on Web and Social Media, 2015.
[6] Christopher Bishop. Pattern Recognition and Machine Learning. Springer, London. 2006.
[7] Alexandra Chouldechova, “Fair prediction with disparate impact: A study of bias in recidivism prediction instruments.” In Conference on Fairness, Accountability, and Transparency (FAT) in Machine Learning, 2017.
[8] Antony Nicol, Chris Casey, Stuart MacFarlane, “Children are ready for speech technology - but is the technology ready for them?” Department of Computing, University of Central Lancashire, 2002.
[9] James A.Rodger, Parag C.Pendharkar, “A field study of the impact of gender and user's technical experience on the performance of voice-activated medical tracking application.” Article in International Journal of Human-Computer Studies 60(5-6):529-544, 2003.
[10] Amit Datta, Michael Carl Tschantz, Anupam Datta, “Automated experiments on ad privacy settings.” Proceedings on Privacy Enhancing Technologies. 2015(1): 92-112. 2015.
[11] Sweeney, L. “Discrimination in online ad delivery.” ACM Queue 11(3): 10, 2013.
[12] Joy Buolamwini, Timnit Gebru, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification.” Proceedings of the 1st Conference on Fairness, Accountability and Transparency (FAT), PMLR 81:77-91, 2018.
[13] World Economic Forum, “Global Gender Gap Report 2017.” Page 30. 2017.
[14] Rachel Rudinger et al. “Gender Bias in Coreference Resolution.” In Proceedings of NAACL-HLT. arXiv preprint arXiv: 1804.09301, 2018.
[15] Trond Petersen, and Laurie A. Morgan, “Separate and Unequal: Occupation-Establishment Sex Segregation and the Gender Wage Gap.” American Journal of Sociology, 101(2): 329-365, 1995.
[16] E. Sapir. Selected Writings in Language, Culture and Personality, volume 342. University of California Press, 1985.
[17] J. Holmes and M. Meyerhoff. The Handbook of Language and Gender, volume 25. John Wiley & Sons, 2008.
[18] Greenwald, A. G., McGhee, D. E., and Schwartz, J. L. “Measuring individual differences in implicit cognition: the implicit association test.” Journal of Personality and Social Psychology, 74(6):1464, 1998.
[19] Kiefer, A. K. and Sekaquaptewa, D. “Implicit stereotypes and women’s math performance: How implicit gender-math stereotypes influence women’s susceptibility to stereotype threat.” Journal of Experimental Social Psychology, 43(5):825-832. 2007.
[20] Nosek, B. A., Banaji, M., and Greenwald, A. G. “Harvesting implicit group attitudes and beliefs from a demonstration web site.” Journal of Group Dynamics: Theory, Research, and Practice, 6(1):101, 2002.
[21] Nosek, B. A., Banaji, M. R., and Greenwald, A. G. “Math= male, me= female, therefore math not = me.” Journal of Personality and Social Psychology, 83(1):44, 2002.
[22] Hinton, G. E., “Learning distributed representations of concepts.”, In R. G. M. Morris (Ed.), Parallel distributed processing: Implications for psychology and neurobiology, pp. 46-61, 1989. (Reprinted from 'Proceedings of the Eighth Annual Conference of the Cognitive Science Society," Amherst, Massachusetts: Erlbaum Associates, 1986)
[23] Yoshua Bengio, Réjean Ducharme, Pascal Vincent, Christian Jauvin; “A Neural Probabilistic Language Model.” Journal of Machine Learning Research (JMLR). 3(Feb):1137-1155, 2003.
[24] T. Mikolov, K. Chen, G. Corrado, and J. Dean. “Efficient Estimation of Word Representations in Vector Space.” Proceeding of International Conference on Learning Representations (ICLR), 2013.
[25] T. Mikolov, I. Sutskever, K. Chen, G. Corrado, and J. Dean. “Distributed Representations of Words and Phrases and their Compositionality.” Proceeding
NIPS'13 Proceedings of the 26th International Conference on Neural Information Processing Systems (NIPS), Volume 2, Pages 3111-3119, 2013.
[26] A. Mnih and K. Kavukcuoglu, “Learning word embeddings efficiently with noise-contrastive estimation.” Proceeding of Advances in Neural Information Processing Systems (NIPS), pp. 2265-2273, 2013.
[27] Jeffrey Pennington, Richard Socher, Christopher D. Manning, “GloVe: Global Vectors for Word Representation.” In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014.
[28] R Collobert et. al. “Natural Language Processing (Almost) from Scratch.” The Journal of Machine Learning Research (JMLR), Volume 12, Pages 2493-2537, Feb. 2011.
[29] Alex Graves et. al. "A Novel Connectionist System for Improved Unconstrained Handwriting Recognition". IEEE Transactions on Pattern Analysis and Machine Intelligence. 31 (5). 2009.
[30] Sak, H & Senior, Andrew & Beaufays, F. “Long short-term memory recurrent neural network architectures for large scale acoustic modeling.” Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, 338-342, 2014.
[31] J. Zhao et al. “Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraints.” On Conference of Empirical Methods in Natural Language Processing (EMNLP, Best long paper), 2017.
[32] R. Pascanu, T. Mikolov, and Y. Bengio, "On the difficulty of training recurrent
neural networks." Conference on International Conference on Machine Learning (ICML), pp.1310-1318, 2013.
[33] S. Hochreiter and J. Schmidhuber, "Long short-term memory." Neural computation, vol. 9, no. 8, pp. 1735-1780, 1997.
[34] I. Sutskever, O. Vinyals, and Q. V. Le, "Sequence to sequence learning with neural networks." In Advances in neural information processing systems (NIPS), pp. 3104-3112, 2014.
[35] D. Kingma and J. Ba, "Adam: A method for stochastic optimization." arXiv
preprint arXiv:1412.6980, 2014.
[36] Nitish Srivastava, Geoffrey Hinton , Alex Krizhevsky , Ilya Sutskever, Ruslan Salakhutdinov, “Dropout: A Simple Way to Prevent Neural Networks from Overfitting.” In Journal of Machine Learning Research (JMLR), 2014
[37] Adam Paszke et. al. “Automatic differentiation in PyTorch.” In Conference on Neural Information Processing Systems (NIPS), 2017.
[38] D. Bahdanau, K. Cho, and Y. Bengio, "Neural machine translation by jointly
learning to align and translate," arXiv preprint arXiv:1409.0473, 2014.
[39] Cho et al. . “Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation.” arXiv preprint arXiv:1406.1708, 2014.
[40] Vinyals et. al., “A Neural Conversational Model.” In Journal of Computer Science, 2015.
[41] Minqing Hu and Bing Liu. "Mining and summarizing customer reviews." Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD-2004, full paper), Seattle, Washington, USA, Aug 22-25, 2004.
[42] Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic Regularities in Continuous Space Word Representations. In Proceedings of NAACL-HLT, 2013.
[43] Eneko Agirre, Enrique Alfonseca, Keith Hall, Jana Kravalova, Marius Pasca, Aitor Soroa, “A Study on Similarity and Relatedness Using Distributional and WordNet-based Approaches,” In Proceedings of NAACL-HLT, 2009.
[44] H. Rubenstein and J. B. Goodenough. “Contextual correlates of synonymy. “ Communications of the ACM, 8(10):627–633, 1965.
[45] Alex Krizhevsky et. al., “ImageNet Classification with Deep Convolutional Neural Networks.” In Conference on Neural Information Processing Systems (NIPS), 2012.
[46] D Silver, A Huang, et al. “Mastering the game of Go with deep neural networks and tree search.” Nature, 529 (7587):484-489, 2016.
[47] 敖恆宇et al. “Family or work : work-family conflict of executives in Hong Kong.” Chinese University of Hong Kong, p156, 1994.