Ya-Ling+Wang++Nan+Mu++Zhe+Fan

Abstract： Deep learning is a new way of multi-layer neural network learning algorithm， which is an unsupervised learning process. The traditional personnel resume information extraction is a supervised learning process. In the process of feature extraction， it is dependent on the amount of manual annotation， paying high cost but low efficiency. This paper presents a model of personnel resume information extraction based on deep learning， and uses the technology of deep learning to study the representation of data characteristics. The process of information extraction is improved by unsupervised learning instead of manual design. Experimental results show that the model is suitable for intelligent extracting personnel resume， and it saves lots of manual annotation effort to improve the efficiency of information extraction.

Keywords： Deep learning； information extraction； Neural Networks； Unsupervised Learning； Auto encoder

1. Introduction

With the development of society， business demand for talent is also increasing. But the face of massive online resume information， it is a challenging task to quickly find companies recruit staffs from these matched resumes. In order to better deal with the relationship between the entity and the context in the resume， you need to choose the suitable characteristics for the task. The appropriate characteristics have clear guidelines for mission objectives， and have a certain degree of improvement for task effect. The characteristics of the word level are a relatively basic unit in Natural Language Processing （NLP）， which compose a lot of semantic and high-level representation， and form a more advanced concept. The traditional resume information extraction process is supervised learning， which depends on a lot of artificial tagging data， so the systems based on supervised need to pay high cost but low efficiency. This paper mainly research on feature extraction task and propose a scheme that the personnel resume feature extraction based on deep learning.

2. Related work

2.1 resume information extraction

As an orientation of information extraction， the personnel resume information extraction has a high commercial value and practical value in the field of application. It can help the large firm to build the talent pool effectively. Ciravegna and Laveli announced the results of a resume information extraction， which use an information extraction tool kit （LP2） to learn the rules of information extraction in English resumes. But it obviously can not meet the needs of real life because the extraction accuracy and range are not enough. In order to solve the problem of address clean， Borkar developed a nested information extraction model， which uses Hidden Markov Model （HMM） to divide the unstructured warehouse address into structured records， and to remove redundant address. On the basis of the above work， Yu Kun puts forward a method based on double-layer cascade text classification， which use for the resume information automatic extraction. This method inherits and extends the nested Borkar&apos；s model， and the model is applied in the field of information extraction， and then a waterfall type information extraction framework is proposed for the structure of resume text.

2.2 Deep Learning

The concept of Deep Learning （DL） is derived from the study of artificial neural network， which named relative to the traditional shallow layer machine learning. Multi-layer Perceptron （MLP） is a feed-forward neural network thermal model， and also belongs to a kind of deep learning structure. Compared with the shallow learning， deep learning uses a distributed feature representation which is a major step forward. In the traditional shallow learning， the representation of the sample uses the form of counting. But the DL is the combination of the underlying features to form a more abstract representation of the higher level. These distributed features are obtained by calculating the multi-hidden layers of the neural network structure layer by layer. In depth study， there are some problems in the local minimum. In the process of optimization， due to the emergence of local minimum， the optimization calculation is likely to terminate before finding the global minimum. This is also the main reason for the difficulty of deep learning in the training. Although deep learning has such a difficult， it reaches a high succeed， one of the advantages it that it has a multi-layers structure， and between these structures are non-linear mapping， which makes DL can be good to complete complex functions approximation. AlphaGo developed by Google&apos；s DeepMind Company whose main principle is deep learning. Overall， Chinese natural language processing based on the deep learning research in this field is still in its infancy， most of the relevant work being implemented and less published literature. So， it has great potential for development and application value.

3. The construction of topic dictionary and topic model

In the data of personnel resume， the effect just based on the characteristics of the word to extract information has some limitations. Further optimization is needed to improve the effect. Word feature is a relatively low-level feature in natural language. By the analysis of the structure and principle of multi-layers automatic coding in deep learning， we can learn that through the deep structure of the neural network can integrate and combine the word level features to form a higher level concept. This hierarchy is also the same as the things level. Therefore， this article will use the unsupervised deep learning to re-expressed personnel resume text features.

4. Experimental results and analysis

4.1 Evaluation Criteria

In the area of information extraction， the general use three indicators to evaluate a method， namely： accuracy rate， recall rate and F-measure. As shown in Fig.6， the collection A represents the correct result， and the B represents the identified result set.

4.2 Experimental results and analysis

Training corpus includes： The 1000 sentence corpus of personnel information for named entity recognition， which most of the sentences contain at least one named entity. The entity is given priority to with person names， place names， organization names. Randomly selected 500 sentences from the baidu encyclopedia contains the information such as education， work experience， the paper presented the sentence extraction with events.

Test corpus was randomly chosen 300 sentence from the Internet， most of the events contained entity （education experience， work experience， published papers）.

5. Summary

In this paper， we introduce the feature learning techniques in depth learning methods， and use the feature learning to replace the manual annotation process in the process of personnel information extracting. But research on deep learning in personnel resume extraction is not very mature， there are many problems need to be further discussed. For example， based on information extracted limitations， the template used in different areas is different， so we have to according to the characteristics to improved model. In the next step of the work need to improve the model， making the system for different tasks in different areas has better adaptability.

References

[1]A. Graves， A. Mohamed， and G. Hinton. Speech recognition with deep recurrent neural networks. In Proceedings of International Conference on Acoustics Speech and Signal Processing （ICASSP）. 2013.

[2]Turney D.P. Learning Algorithms for Keyphrase Extraction[J]. Information Retrieval， 2000， 2（4）：303-336.

[3]Y. Bengio. Deep learning of representations： Looking forward. In Statistical Language and Speech Processing， pages 1–37. Springer， 2013.

[4]Liu H， Taniguchi T. Feature Extraction and Pattern Recognition for Human Motion by a Deep Sparse Autoencoder[C]. Computer and Information Technology （CIT）， 2014 IEEE International Conference on. IEEE， 2014：173-181.

[5]Deco G， Stetter M， Szabo M. Method for computer-aided learning of a neural network and neural network： US， US 8423490 B2[P]. 2013.

[6]Dahl G E，Yu D，Deng L，et al. Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition[J].Audio， Speech， and Language Processing， IEEE Transaction on， 2012， 20（1）：30-42.