臺灣博碩士論文加值系統

近年來台灣社會飛速老化，將來可能面臨醫療與老人照護方面人力資源不足的情況，雖然面臨著困境，但這也代表著一個新的市場，隨著科技進步與硬體的提升，現今人工智慧隨處可見，並且逐步取代人類部分簡單的工作，因此希望透過自然語言處理(Natural Language Processing)技術，能夠幫助分攤人力資源的不足。本研究著重於能夠讓機器代替人類與老人進行對話，以滿足老人心靈的孤單感，因此流暢的對答與語意通順對於本研究非常重要。一個成功的長期對話需要考慮到整個對話串的前後文，對於語意相同的問題回覆一致的答案，因此需要從上下文中合併更多的訊息，然而機器本身並不懂語言的意義，機器只專注於數學上的計算，人類使用的自然語言由於一語多義的特性與複雜度，使得語言並不能和數字相提並論，生成式對話的隨機性也將影響回答結果，因此在長期對話中該如何讓機器保持一貫的人格將會是困難點之一。本研究使用簡短對話集與健康領域文章當作訓練語料，簡短對話集作為，本研究使用語意相似度與困惑度(Perplexity)進行對話的篩選與評估，並在篩選過程中利用語意相似度比對選擇最佳的回答句，在對話進行中使用語意相似度與困惑度評估對話串內容是否連貫進而進行修改，以保證對話能夠在一定的限制下進行。

In recent years, Taiwan's society has been rapidly aging. In the future, it may face a shortage of human resources in medical and elderly care. Although it faces difficulties, it also represents a new market. With the advancement of science and technology and the improvement of hardware, artificial intelligence is everywhere today. It can be seen, and gradually replaces some of the simple tasks of human beings. Therefore, it is hoped that the natural language processing technology can help distribute the shortage of human resources. This study focuses on enabling machines to replace humans in conversations with the elderly to satisfy the loneliness of the elderly's mind. Therefore, smooth response and semantic coherence are very important for this research. A successful long-term conversation needs to take into account the context of the entire dialogue string, and reply to the same question with the same semantics. Therefore, it is necessary to incorporate more information from the context. However, the machine itself does not understand the meaning of language, and the machine only focuses on Mathematical calculation, due to the polysemy and complexity of natural language used by humans, language cannot be compared with numbers, and the randomness of generative dialogue will also affect the answer results, so how to keep the machine in long-term dialogue? Consistent personality will be one of the difficulties. This study uses a short dialogue set and articles in the field of health as training corpus, and a short dialogue set as the training corpus. This research uses semantic similarity and perplexity to screen and evaluate dialogue, and use the semantic similarity ratio in the screening process. For the selection of the best answer sentences, semantic similarity and confusion are used to evaluate whether the content of the dialogue string is coherent and then modified to ensure that the dialogue can be carried out under certain restrictions.

摘要 i
ABSTRACT ii
目錄 iii
表目錄 v
圖目錄 vii
壹、緒論 1
1.1研究背景與動機 1
1.2研究目的 2
1.3論文架構 3
貳、文獻探討 4
2.1自然語言生成(Natural Language Generation, NLG) 4
2.2老人照護(Elderly Care) 4
2.3自注意力機制(Self-Attention) 5
2.4 GPT語言模型 5
2.5 Bert語言模型 6
參、研究架構與方法 8
3.1研究架構 8
3.2研究流程 8
肆、實驗結果 12
4.1 一般對話 12
4.2 附加詞 12
4.3語句偵測與替換 18
4.4句子篩選流程 24
4.4.1 回應句篩選流程 24
4.4.2 替換句篩選流程 25
五、結論與未來展望 27
參考文獻 29
附錄 30
附錄一 30
附錄二 31
附錄三 33

Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever. (2018). Improving Language Understanding.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin. (2017). Attention Is All You Need. Neural Information Processing Systems.
Shirina Hannan, Ian James Norman, SALLY J. REDFERN. (2001). Care Work and Quality of Care for Older People: A Review of the Research Literature. Reviews in Clinical Gerontology.
Thibault Sellam, Dipanjan Das, Ankur P. Parikh. (2020). BLEURT: Learning Robust Metrics for Text Generation.
Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocký. (2010). Recurrent neural network based language model. INTERSPEECH 2010, 11th Annual Conference of the International Speech Communication Association.
Ye Jia, Yu Zhang, Ron J. Weiss, Quan Wang, Jonathan Shen, Fei Ren, Zhifeng Chen, Patrick Nguyen, Ruoming Pang, Ignacio Lopez Moreno, Yonghui Wu. (2018). Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis. Advances in Neural Information Processing Systems.
自然語言處理的研究與發展. (2013). 李生. 燕山大學學報.