基于特征融合的中文简历解析方法研究
|
摘要
针对基于规则和统计的传统中文简历解析方法效率低、成本高、泛化能力差的缺点,提出一种基于特征融合的中文简历解析方法,即级联Word2Vec生成的词向量和用BLSTM(Bidirectional Long Short-Term Memory)建模字序列生成的词向量,然后再结合BLSTM和CRF(Conditional Random Fields)对中文简历进行解析(BLSTM-CRF)。为了提高中文简历解析的效率,级联包含字序列信息的词向量和用Word2Vec生成的词向量,融合成一个新的词向量表示;再由BLSTM强大的学习能力融合词的上下文信息,输出所有可能标签序列的分值给CRF层;再由CRF引入标签之间约束关系求解最优序列。利用梯度下降算法训练神经网络,使用预先训练的词向量和Dropout优化神经网络,最终完成对中文简历的解析工作。实验结果表明,所提的特征融合方法优于传统的简历解析方法。
It's typical for the Chinese resume analysis to apply the rule-based and statistical-based methods, suffering from the low efficiency, high cost and poor generalization ability. This paper proposes a Chinese resume analysis method based on feature fusion model. The concatenation of the word vectors generated by Word2 Vec and the word representation is generated from BLSTM neural network, then the text resume is analyzed by intergrating the BLSTM and CRF model(BLSTM-CRF). In order to improve the efficiency of Chinese resume resolution, the two vectors are concatenated into a new word representation. Furthermore, the BLSTM layer is used to fuse the contextual information of the words to be marked, and then the values of all possible tag sequences are exported to the CRF layer. Finally, according to the constraints of the front and rear labels, the CRF is utilized to obtain the optimal labeling sequence. All of the neural networks are trained by the gradient descent algorithm and are optimized by the pretrained word embeddings and Dropout. The experimental results show that the feature fusion method is superior to the traditional resume analysis schemes.
It's typical for the Chinese resume analysis to apply the rule-based and statistical-based methods, suffering from the low efficiency, high cost and poor generalization ability. This paper proposes a Chinese resume analysis method based on feature fusion model. The concatenation of the word vectors generated by Word2 Vec and the word representation is generated from BLSTM neural network, then the text resume is analyzed by intergrating the BLSTM and CRF model(BLSTM-CRF). In order to improve the efficiency of Chinese resume resolution, the two vectors are concatenated into a new word representation. Furthermore, the BLSTM layer is used to fuse the contextual information of the words to be marked, and then the values of all possible tag sequences are exported to the CRF layer. Finally, according to the constraints of the front and rear labels, the CRF is utilized to obtain the optimal labeling sequence. All of the neural networks are trained by the gradient descent algorithm and are optimized by the pretrained word embeddings and Dropout. The experimental results show that the feature fusion method is superior to the traditional resume analysis schemes.
引文
[1]陈川波.基于半结构化文本信息抽取的简历识别系统[D].北京:北京邮电大学,2008.
[2]江志祥.智能简历解析系统的研究与实现[D].北京:北京邮电大学,2009.
[3]杨永贵.中文信息抽取关键技术研究与实现[D].北京:北京邮电大学,2008.
[4]Lee L M,Lee J C.A study on high-order hidden Markov models and applications to speech recognition[J].Advances in Applied Artificial Intelligence,2006,4031(34):682-690.
[5]Mc Callum A,Freitag D,Pereira F C N.Maximum entropy Markov models for information extraction and segmentation[C]//Proceedings of the 18th International Conference on Machine Learning,2000:591-598.
[6]Sutton C,McCallum A.An introduction to conditional random fields[J].Foundations and Trends in Machine Learning,2012,4(4):367-373.
[7]Mesnil G,Dauphin Y,Yao K,et al.Using recurrent neural networks for slot filling in spoken language understanding[J].IEEE/ACM Transactions on Audio,Speech and Language Processing,2015,23(3):530-539.
[8]Plank B,Sogaard A,Goldberg Y,et al.Multilingual part-ofspeech tagging with bidirectional long short-term memory models and auxiliary loss[J].Meeting of the Association for Computational Linguistics,2016,10(2):412-418.
[9]Lu J L,Kato M P,Yamamoto T,et al.Entity identification on microblogs by CRF model with adaptive dependency[C]//IEEE/Wic/ACM International Conference on Web Intelligence and Intelligent Agent Technology,2016:333-340.
[10]Hinton G E,Srivastava N,Krizhevsky A,et al.Improving neural networks by preventing co-adaptation of feature detectors[J].Computer Science,2012,3(4):212-223.
[11]Mikolov T,Sutskever I,Chen K,et al.Distributed representations of words and phrases and their compositionality[J].Neural Information Processing Systems,2013,5(4):3111-3119.
[12]Pennington J,Socher R,Manning C.Glove:global vectors for word representation[C]//Conference on Empirical Methods in Natural Language Processing,2014:1532-1543.
[13]Zhang Xiang,Zhao Junbo,Le Cun Yann.Character-level convolutional networks for text classification[J].Advances in Neural Information Processing Systems,2015,12(2):649-657.
[14]Lee J H,Delbruck T,Pfeiffer M.Training deep spiking neural networks using backpropagation[J].Frontiers in Neuroscience,2016,10:662-668.
[15]Khirirat S,Feyzmahdavian H R,Johansson M.Mini-batch gradient descent:faster convergence under data sparsity[C]//IEEE Conference on Decision and Control,2017:2880-2887.
[16]李航.统计学习方法[M].北京:清华大学出版社,2012:14-15.
[2]江志祥.智能简历解析系统的研究与实现[D].北京:北京邮电大学,2009.
[3]杨永贵.中文信息抽取关键技术研究与实现[D].北京:北京邮电大学,2008.
[4]Lee L M,Lee J C.A study on high-order hidden Markov models and applications to speech recognition[J].Advances in Applied Artificial Intelligence,2006,4031(34):682-690.
[5]Mc Callum A,Freitag D,Pereira F C N.Maximum entropy Markov models for information extraction and segmentation[C]//Proceedings of the 18th International Conference on Machine Learning,2000:591-598.
[6]Sutton C,McCallum A.An introduction to conditional random fields[J].Foundations and Trends in Machine Learning,2012,4(4):367-373.
[7]Mesnil G,Dauphin Y,Yao K,et al.Using recurrent neural networks for slot filling in spoken language understanding[J].IEEE/ACM Transactions on Audio,Speech and Language Processing,2015,23(3):530-539.
[8]Plank B,Sogaard A,Goldberg Y,et al.Multilingual part-ofspeech tagging with bidirectional long short-term memory models and auxiliary loss[J].Meeting of the Association for Computational Linguistics,2016,10(2):412-418.
[9]Lu J L,Kato M P,Yamamoto T,et al.Entity identification on microblogs by CRF model with adaptive dependency[C]//IEEE/Wic/ACM International Conference on Web Intelligence and Intelligent Agent Technology,2016:333-340.
[10]Hinton G E,Srivastava N,Krizhevsky A,et al.Improving neural networks by preventing co-adaptation of feature detectors[J].Computer Science,2012,3(4):212-223.
[11]Mikolov T,Sutskever I,Chen K,et al.Distributed representations of words and phrases and their compositionality[J].Neural Information Processing Systems,2013,5(4):3111-3119.
[12]Pennington J,Socher R,Manning C.Glove:global vectors for word representation[C]//Conference on Empirical Methods in Natural Language Processing,2014:1532-1543.
[13]Zhang Xiang,Zhao Junbo,Le Cun Yann.Character-level convolutional networks for text classification[J].Advances in Neural Information Processing Systems,2015,12(2):649-657.
[14]Lee J H,Delbruck T,Pfeiffer M.Training deep spiking neural networks using backpropagation[J].Frontiers in Neuroscience,2016,10:662-668.
[15]Khirirat S,Feyzmahdavian H R,Johansson M.Mini-batch gradient descent:faster convergence under data sparsity[C]//IEEE Conference on Decision and Control,2017:2880-2887.
[16]李航.统计学习方法[M].北京:清华大学出版社,2012:14-15.