| 引用本文: | 金志刚,韩玥,朱琦.一种结合深度学习和集成学习的情感分析模型[J].哈尔滨工业大学学报,2018,50(11):32.DOI:10.11918/j.issn.0367-6234.201709078 |
| JIN Zhigang,HAN Yue,ZHU Qi.A sentiment analysis model with the combination of deep learning and ensemble learning[J].Journal of Harbin Institute of Technology,2018,50(11):32.DOI:10.11918/j.issn.0367-6234.201709078 |
|
| 摘要: |
| 随着社交媒体的不断发展, 用户评价已成为网络决策的关键因素.为了准确分析社交媒体用户评价的情感倾向性, 更好地推进舆情分析、推荐算法等工作, 本文通过对Bi-LSTM模型和Bagging算法的改进, 提出了一种新的情感分析模型—Bi-LSTMM-B模型.该模型的特点在于将深度学习模型可提取抽象特征的优势和集成学习多分类器共同决策的思想相结合.一方面在Bi-LSTM模型的基础上引入Maxout神经元, 构建Bi-LSTMM模型, 解决随机梯度下降算法中存在的梯度弥散问题, 更好地优化训练过程.另一方面, 模型基于Bagging算法训练多个情感分类器, 根据分类器性能优劣利用袋外数据为每个分类器分配指定类别的权重, 并提出相应的改进投票策略, 增强了模型的泛化能力.实验结果表明:本文提出的Bi-LSTMM-B模型相比于传统的LSTM模型准确率提高12.08%, 其中Maxout神经元的引入对情感分析准确率有8.28%的相对改善效果, 改进后的投票策略对准确率有4.06%的相对改善效果, 并在召回率和F值两项指标上均优于其他对比模型.由此证明, 深度学习模型和集成学习思想相结合可提高情感分析的准确率, 并具有一定的研究价值
|
| 关键词: 情感倾向性分析 深度学习 集成学习 Bi-LSTM模型 Maxout神经元 Bagging算法 |
| DOI：10.11918/j.issn.0367-6234.201709078 |
| 分类号:TP391 |
| 文献标识码:A |
| 基金项目:国家自然科学基金项目(71502125) |
|
| A sentiment analysis model with the combination of deep learning and ensemble learning |
|
JIN Zhigang,HAN Yue,ZHU Qi
|
|
(School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China)
|
| Abstract: |
| With the development of social media, users' evaluations have become a key factor in network decision-making.Owing to the necessity of making a more accurate analysis on the emotional tendency of social media users' evaluations as well as promoting public opinion analysis and recommendation algorithms, a sentiment analysis model called Bi-LSTMM-B (Bi-directional Long Short Term Memory Model with Maxout neurons in Bagging algorithm) is proposed.With the feature of combining deep learning model and the idea of ensemble learning, the model improves the Bi-LSTM model and the Bagging algorithm.On the one hand, the Bi-LSTMM model introduces the Maxout neural into the Bi-LSTM model to solve the vanishing gradient problem during the stochastic gradient descent training and optimize the training process.On the other hand, multiple emotional classifiers were trained at the foundation of the Bagging algorithm.The out of bag data assigns the weight for each classifier on specified category according to their performance.Hence the voting strategy is improved to enhance the generalization ability of the model.The experimental results indicate that the accuracy of the Bi-LSTMM-B model is improved by 12.08% compared to the traditional LSTM model.It is also superior to other contrast models in the recall rate and F value.Therein, the introduction of Maxout neurons has a relative improvement effect of 8.28% on the accuracy of sentiment analysis, while the improved voting strategy accounts for 4.06% on the accuracy.Thus, it proves that combining deep learning and ensemble learning contributes to the improvement of the accuracy of sentiment analysis, which shows some value in research.
|
| Key words: sentiment analysis deep learning ensemble learning Bi-LSTM model Maxout neural Bagging algorithm |
