第四章 實驗結果與討論
第三節 分析與討論
0.99 0.9 0.94 850.27
第三節 分析與討論
表 4.3.1 整理上述各個實驗設計的效能。可以看出比起單純使用全連結的類 神經網路，不管使用 CNN encoder 特徵擷取或 LSTM encoder 優化 word2vec 都
讓效果有所提升；LSTM 架構的損失值與損失值間的差都是最小的，而 FCNN 架
62
構的損失值與損失值間的差是最大的。使用雙 encoder 架構下，只有雙 CNN 架 構的結果有變好，LSTM 則無；代表雙 CNN 架構進一步擷取特徵對建議探勘是 有用的，而雙 LSTM 作為優化 word2vec 的功用使用一次就足夠；使用順逆向 LSTM (Bi-LSTM)改善 word2vec 上下文距離短的缺點之想法，在建議探勘中效果 沒有顯著提升，Recall 甚至下降，可能原因是只要有建議行為之句子就被標記為
63 份 : (1) my recommendation should really read yes and no. (2) highly recommend for
all around the house use. (3) they 'll learn one day that having too many options on the store counters will run consumers to a product with less options.，本研究發現，不論
建議成份出現在主詞或動詞都可以識別出建議類。而下面三句標記為建議類，但
64
模型識別為非建議類，底線加粗體是建議行為的部份，斜體是誤判項 : (1) looks
like a nice answer to organize presentation papers. (2) practical and can contained
sorted documents in one folder. (3) a laminated version would be awesome i prefer a single-color pack，本研究發現，一個句子如果包含形容詞補語或是感官動詞，就
算句子裡有建議行為的部份，也會被識別為非建議類。
65
表4.3.1 實驗之效能清單
Precision Recall F1 Training time
(s)
66
67
68
或許可以讓模型的效能更上一層。
全連結網路、卷積神經網路或長短期記憶網路每種網路都有各自的特色。當
中可以調整的參數非常多，例如：全連結層可以調整隱藏層數目、每層的神經元 個數或激活函數等；卷積神經網路可以調整過濾器個數與大小、卷積移動的策略、
池化策略或激活函數等；長短期記憶網路可以調整網路輸出的維度或激活函數等。
未來可以嘗試的組合有很多種，或許有一種網路組合與參數可以讓原本不突出的 詞表達式與網路模型得到更好的效能。
建議分類可以找出明顯建議的評論，提供消費者購買前的參考依據，或是提
供給業者迭代設計的建議。此分類並無提供訊息是給與消費者或業者。一個可以 進一步延伸的主題是根據分類出的評論再分出此評論是給消費者或業者。
69
參考文獻
Brun, C., & Hagege, C. (2013). Suggestion Mining: Detecting Suggestions for
Improvement in Users' Comments. Research in Computing Science, 70(79.7179), 5379-62.
Bengio, Y., Ducharme, R., Vincent, P., & Jauvin, C. (2003). A neural probabilistic language model. Journal of machine learning research, 3(Feb), 1137-1155.
Dong, L., Wei, F., Duan, Y., Liu, X., Zhou, M., & Xu, K. (2013, June). The automated acquisition of suggestions from tweets. In Twenty-Seventh AAAI Conference on Artificial Intelligence.
Fernández, A. M., Esuli, A., & Sebastiani, F. (2016). Distributional Correspondence Indexing for Cross-Lingual and Cross-Domain Sentiment Classification. Journal of artificial intelligence research, 55, 131-163.
Golchha, H., Gupta, D., Ekbal, A., & Bhattacharyya, P. (2018). Helping each Other: A Framework for Customer-to-Customer Suggestion Mining using a Semi-supervised Deep Neural Network. arXiv preprint arXiv:1811.00379.
Kim, Y. (2014). Convolutional neural networks for sentence classification. arXiv preprint arXiv:1408.5882.
Liu, B. (2012). Sentiment analysis and opinion mining. Synthesis lectures on human language technologies, 5(1), 1-167.
70
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013a). Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems (pp. 3111-3119).
Mikolov, Tomas, et al. "Efficient estimation of word representations in vector space."
arXiv preprint arXiv:1301.3781 (2013b).
Negi, S., & Buitelaar, P. (2015). Towards the extraction of customer-to-customer suggestions from reviews. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (pp. 2159-2167).
Negi, S., Asooja, K., Mehrotra, S., & Buitelaar, P. (2016). A study of suggestions in opinionated texts and their automatic detection. In Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics (pp. 170-178).
Pei, J., Han, J., Mortazavi-Asl, B., Wang, J., Pinto, H., Chen, Q., ... & Hsu, M. C.
(2004). Mining sequential patterns by pattern-growth: The prefixspan approach.
IEEE Transactions on knowledge and data engineering, 16(11), 1424-1440.
Pennington, J., Socher, R., & Manning, C. (2014). Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) (pp. 1532-1543).
Ramanand, J., Bhavsar, K., & Pedanekar, N. (2010, June). Wishful thinking: finding suggestions and'buy'wishes from product reviews. In Proceedings of the
71
NAACL HLT 2010 workshop on computational approaches to analysis and generation of emotion in text (pp. 54-61). Association for Computational Linguistics.
Rendle, Steffen. "Factorization machines with libfm." ACM Transactions on Intelligent Systems and Technology (TIST) 3.3 (2012): 57.