本研究利用注意力機制模型偵測財務文章之風險事件及抽取潛在金融犯罪名單，建構自動化模型以降低人力標記成本及提升預測速度。我們分析不同模型架構及訓練方法之優缺點，並比較傳統神經網路方法與 Transformer Based 模型的差異。模型架構分為兩階段，第一階段判斷目標文章是否包含金融風險事件，而第二階段則在這些文章中抽取高危險的名單。我們提出聯合訓練方法同時訓練兩階段的模型，透過實驗證明可在不損失正確性的情況提升訓練及預測速度，並得以提升模型穩定性。我們亦針對注意力機制模型內部的 Attention Weight 做視覺化分析，顯示模型能在不提供標注的情況自動關注金融風險詞彙。另外我們針對缺乏風險人名標記的訓練資料之情況，利用以上 Attention Weight 分析設計特殊的規則，達到一定程度的效果提升。最後我們額外在一個 Wikipedia 上的英文資料集做測試，說明此研究結果亦可應用於不同領域及不同語言的任務。

This thesis uses transformer-based models to detect risk events from financial articles and extract potential financial criminals. With such automated models, we can reduce human costs on labeling and increase prediction performance. In this thesis, we analyze the advantages and disadvantages of different approaches and compare the differences between traditional neural networks and
Transformer-based models. The proposed method contains two stages: the first stage determines whether the target news contains financial risk events, and the second stage extracts high-risk entities from the news. We propose a
joint-training method to train these two stages at the same time. Experimental results show that the proposed joint-training method improves prediction accuracy and enhances the stability of the training process. We also visualize
the attention weights of the attention mechanism model, showing that the model automatically pays attention to financial risk vocabularies without providing annotations. In addition, we use the above attention weight scheme to design special rules, achieving a certain degree of effect improvement for the case that lacks risk-name-annotation. Finally, further experiments conducted on a dataset from English Wikipedia confirm that the proposed method can also apply to different domains and languages.

致謝 i
摘要 ii
Abstract iii
目錄 iv
圖目錄 vi
表目錄 vii
第一章 緒論 1
1.1 前言 1
1.2 研究目的與貢獻 3
第二章 相關文獻探討 4
2.1 自然語言中的文字表示法 4
2.2 中文斷詞 5
2.3 循環神經網路 5
2.4 Transformer-Based模型 5
2.4.1 注意力機制模型 5
2.4.2 Transformer 6
2.4.3 BERT 6
2.5 命名實體識別 7
第三章 研究方法 9
3.1 問題定義 9
3.2 模型簡介 10
3.2.1 新聞文本分類任務 — CLASS任務 12
3.2.2 實體辨識任務 — NER任務 13
3.2.3 聯合訓練 — Joint Training 13
3.3 各種架構之模型實作方式 14
3.3.1 LSTM架構實作 14
3.3.2 Attention架構實作 15
3.3.3 混合使用LSTM與Attention架構 16
3.3.4 BERT架構實作 17
3.4 利用Attention Weight配合通用NER工具之實作 17
第四章 實驗結果 19
4.1 資料說明 19
4.1.1 Wikipedia資料（Wiki資料集） 19
4.1.2 新聞資料（News資料集） 20
4.2 實驗設定 20
4.3 實驗結果分析 21
4.3.1 文章分類結果 21
4.3.2 實體辨識結果 23
4.3.3 目標實體抽取結果 25
4.4 模型訓練參數分析 28
4.4.1 依Positive及Negative資料比例調整Loss權重之分析 28
4.4.2 Joint模型兩任務Loss之權重比較 29
4.5 Attention NER模型學習到之資訊分析 30
4.6 Attention Weight分析 30
第五章 結論 32
參考文獻 34

[1] D. W. Otter, J. R. Medina, and J. K. Kalita, “A survey of the usages of deep learn- ing for natural language processing,” IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 2, pp. 604–624, 2020.
[2] R.Jozefowicz,W.Zaremba,andI.Sutskever,“Anempiricalexplorationofrecurrent network architectures,” in International conference on machine learning. PMLR, 2015, pp. 2342–2350.
[3] S.HochreiterandJ.Schmidhuber,“Longshort-termmemory,”Neuralcomputation, vol. 9, no. 8, pp. 1735–1780, 1997.
[4] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” arXiv preprint arXiv:1706.03762, 2017.
[5] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018.
[6] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al., “Language models are unsupervised multitask learners,” OpenAI blog, vol. 1, no. 8, p. 9, 2019.
[7] K. Potdar, T. S. Pardawala, and C. D. Pai, “A comparative study of categorical vari- able encoding techniques for neural network classifiers,” International journal of computer applications, vol. 175, no. 4, pp. 7–9, 2017.
[8] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of word repre- sentations in vector space,” arXiv preprint arXiv:1301.3781, 2013.
[9] Q. Liu, M. J. Kusner, and P. Blunsom, “A survey on contextual embeddings,” arXiv preprint arXiv:2003.07278, 2020.
[10] J. Li, A. Sun, J. Han, and C. Li, “A survey on deep learning for named entity recog- nition,” IEEE Transactions on Knowledge and Data Engineering, pp. 1–1, 2020.
[11] V.KrishnanandV.Ganapathy,“Namedentityrecognition,”StanfordLectureCS229, 2005.
[12] S. R. Eddy, “Hidden markov models,” Current opinion in structural biology, vol. 6, no. 3, pp. 361–365, 1996.
[13] J. Lafferty, A. McCallum, and F. C. Pereira, “Conditional random fields: Probabilis- tic models for segmenting and labeling sequence data,” 2001.
[14] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high-performance deep learning library,” in Advances in Neural Information Processing Systems 32, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, Eds. Curran Associates, Inc., 2019, pp. 8024–8035. [Online]. Available: http://papers.neurips.cc/paper/9015-pytorch-an- imperative-style-high-performance-deep-learning-library.pdf
[15] T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. L. Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush, “Transformers: State-of-the-art natural language processing,” in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Online: Association for Computational Linguistics, Oct. 2020, pp. 38–45. [Online]. Available: https://www.aclweb.org/anthology/2020.emnlp-demos.6
[16] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.
[17] I.LoshchilovandF.Hutter,“Decoupledweightdecayregularization,”arXivpreprint arXiv:1711.05101, 2017.