建置與評估文字自動生成的情感對話系統

Development and Evaluation of Emotional Conversation System Based on Automated Text Generation

楊德倫(Te-Lun Yang) ； 曾元顯(Yuen-Hsien Tseng)

《教育資料與圖書館學》 57卷3期 (2020/11) Pp. 355-378

https://doi.org/10.6120/JoEMLS.202011_57(3).0048.RS.CM

摘要

本研究藉由2019年中文情緒對話生成（CECG）評比任務所提供約170萬則語料，運用深度學習GPT-2與BERT等技術與工具，實作了具備情感對話的系統，並以CECG提供的測試發文評估其成效。由三位人工判斷的結果，顯示本研究建置發展的系統，與2019年CECG評測最佳團隊的系統有類似的成效水準。而進一步的案例分析發現，對於訓練資料中較普遍的話題，GPT-2的語言建模技術，的確可以生成創新、有趣、完美的回應文句。本研究的主要貢獻為：㈠將情感融入發文字串中做為條件求機率，以便簡潔地依原方式訓練並使用GPT-2；㈡運用BERT來預測回應文句的連貫性以做為排序的依據。雖然這兩項技巧分別源自GPT與BERT的訓練機制，但本研究稍加修改應用於CECG的任務上，獲得了不錯的效果。

關鍵字

對話系統 ； 文字生成 ； 文意理解 ； 深度學習 ； 人工智慧

並列摘要

Based on the corpus provided by the 2019 Chinese Emotional Conversation Generation (CECG) evaluation task, an emotional conversation system is implemented in this paper using deep learning and other technologies such as GPT-2 and BERT. The effectiveness of the system is evaluated based on the test data and criteria provided by CECG. The results based on three human annotators show that the system has a similar effectiveness level with that of the best team participating in the 2019 CECG task. Further case studies reveal that the more post/reply pairs about a topic in the training data, the better the language model of GPT-2 to generate innovative, interesting, and perfect response sentences for that topic. The main contributions of this study are: 1. Integrating emotion into the post string as a condition for computing probability, so as to simply train GPT-2 and make GPT-2 predict in the original way; 2. Applying BERT to predict the coherence of response sentences as a basis for ranking. Although these two techniques are derived from the training mechanisms of GPT and BERT respectively, we have slightly modified them to fit the task of CECG and achieved good results.

並列關鍵字

Conversational system ； Text generation ； Text understanding ； Deep learning ； Artificial intelligence

參考文獻

Binsted, K., Bergen, B., Coulson, S., Nijholt, A., Stock, O., Strapparava, C., Ritchie, G., Manurung, R., Pain, H., Waller, A., & O’Mara, D. (2006). Computational humor. IEEE Intelligent Systems, 21(2), 59-69. https://doi.org/10.1109/MIS.2006.22

Cagan, T., Frank, S. L., & Tsarfaty, R. (2017). Data-driven broad-coverage grammars for opinionated natural language generation (ONLG). In R. Barzilay & M.-Y. Kan (Eds.), Proceedings of the 55th annual meeting of the Association for Computational Linguistics (Vol. 1, pp. 1331-1341). Association for Computational Linguistics. https://doi.org/10.18653/v1/P17-1122

Ghosh, S., Chollet, M., Laksana, E., Morency, L.-P., & Scherer, S. (2017). Affect-LM: A neural language model for customizable affective text generation. In R. Barzilay & M.-Y. Kan (Eds.), Proceedings of the 55th annual meeting of the Association for Computational Linguistics (Vol. 1, pp. 634-642). Association for Computational Linguistics. https://doi.org/10.18653/v1/P17-1059

Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780. https://doi.org/10.1162/neco.1997.9.8.1735

Hornik, K. (1991). Approximation capabilities of multilayer feedforward networks. Neural Networks, 4(2), 251-257. https://doi.org/10.1016/0893-6080(91)90009-T