Computer and Modernization

Question Classification Based on Hybrid Neural Network Model

CHEN Ke-jin1,2,3, XU Guang-luan2,3, GUO Zhi2,3, LIANG Xiao2,3

2018, 0(09): 1. doi:10.3969/j.issn.1006-2475.2018.09.001

Asbtract ( 138 )

References | Related Articles | Metrics

The automatic question answering system gives fast and accurate answers to the questions proposed by the users in natural language, arousing widespread concern in academia and industry. By automatically determining the type of question, question classification task is of great significance to improve the accuracy of the question answering system. Based on the contextual information of the question and answer, combined with the respective advantages of convolutional neural networks and recurrent neural networks, this paper proposes a hybrid deep learning model. In addition, in order to strengthen the representation capacity of the question, this model adopts attention mechanism and enhances the generalization ability of the model. In this paper, we conduct a comparative experiment on 360 QA datasets, results show that this model has improved 1.6%~5.6% compared with the traditional method.

Knowledge Graph Question Answering Based on Multi-granularity Feature Representation

SHEN Cun1,2,3, HUANG Ting-lei2,3, LIANG Xiao2,3

2018, 0(09): 5. doi:10.3969/j.issn.1006-2475.2018.09.002

Asbtract ( 337 )

References | Related Articles | Metrics

Recently, knowledge graph question answering has gradually become the focus of academic and industrial circles. However, traditional methods often have problems of inefficiency and insufficient use of data information. In order to solve the problems above, this paper divides the Chinese knowledge graph question answering into two sub-tasks: entity extraction and property selection. The Bi-LSTM-CRF model is used to identify entities, and a multi-granularity feature representation model is proposed to carry out property selection. The model utilizes character-level and word-level to represent questions and properties and encode them through the encoder. At the same time, it also introduces the one-hot information for the properties. Through the combination of multi-granularity text representations and the similarity calculation of questions and properties, the system finally achieves a 73.96% F1 value on the NLPCC-ICCPOL 2016 KBQA data set, which finishes the knowledge graph question and answer task successfully.

Review of Robot Arm Grasping Behavior Planning

HAN Li-li, WANG Qi-zhi, YANG Yong-gang

2018, 0(09): 11. doi:10.3969/j.issn.1006-2475.2018.09.003

Asbtract ( 248 )

References | Related Articles | Metrics

As the robot’s working environment becomes more and more complex, through exemplary teaching to them, we can make the robot adapt to its environment faster and better accomplish the missions that humans have given. In the robot’s behavior, the robot arm is indispensable in the robot’s work. In recent years, the modeling and control of robotic arm grabbing has become a hot topic in scientific research. The most cutting-edge methods start with behavior acquisition and behavioral representation. There are roughly two methods for behavior acquisition, one is the real-person teaching method, the other is the virtual platform teaching method, and most of the behavioral characterization uses deep learning method. In this paper, the method of data acquisition is used as an example to review the research and development of manipulator modeling and control in recent years.

Chinese Personal Relation Extraction Method Based on Convolutional Neural Network

SI Wen-hao1, JIA Lei-ping2, QI Yin-cheng2

2018, 0(09): 17. doi:10.3969/j.issn.1006-2475.2018.09.004

Asbtract ( 314 )

References | Related Articles | Metrics

Focused on the problem that the features need to be selected manually in personal relation extraction based on machine learning, a Chinese personal relation extraction method based on convolutional neural networks is proposed. The Word2vec model is trained by the Internet Chinese news corpus of Sogou Lab, and the expression of word vector based on distributed representation is obtained, and the transformation of the word vector for the Baidu encyclopedia data set is completed. A Chinese personal relation extraction system based on the classic CNN model is designed. The features are automatically extracted and the personal relation is classified by the CNN model. The accuracy rate reaches to 92.87%, and the average recall rate reaches to 86.92% in extraction of 5 kinds of personal relation. Experimental results show that this method does not need to construct complex features artificially, and it can get a better effect in personal relation extraction.

Information Extraction of Web Pages Based on Support Vector Machine

LIANG Dong, YANG Yong-quan, WEI Zhi-qiang

2018, 0(09): 21. doi:10.3969/j.issn.1006-2475.2018.09.005

Asbtract ( 249 )

References | Related Articles | Metrics

Aiming at the text information extraction of Web pages, this paper presents a method of extracting text information based on support vector machines. This method adopts “come in easily, out strictly” policy. The first step is to traverse the Web DOM tree according to the rules of the Web page structure, and locate an HTML tag that contains both useful and noise information. The second step is to select five important features of the HTML tag with noise information and use SVM to train the sample data. The model can effectively remove the navigation, promotion, copyright and other noise information, and preserve the useful information of Web pages. The method is applied to several commonly used websites. The experimental results show that this method has good effect of extracting texts and noise reduction, and can preserve short texts, such as hyperlinks related to texts that often mistakenly deleted by traditional methods.

An Optimization Method of SVM Parameters Based on Improved QPSO Algorithm

ZHOU Di1,2

2018, 0(09): 27. doi:10.3969/j.issn.1006-2475.2018.09.006

Asbtract ( 141 )

References | Related Articles | Metrics

Because the parameter selection of support vector machine (SVM) has an important influence on the modeling precision and generalization performance, an optimization method of SVM parameters based on improved QPSO algorithm (IQPSO-SVM) is proposed. In the IQPSO-SVM method, the mixed disturbance operator is introduced into QPSO algorithm in order to obtain the average optimization position to construct an improved QPSO (IQPSO) algorithm. Then the IQPSO algorithm with the global optimization ability is used to optimize the penalty coefficient and kernel parameter of SVM model in order to obtain the optimal combination values of parameters and improve the accuracy and solving speed for SVM model. The test functions and UCI data are used to test and verify the effectiveness of the proposed IQPSO-SVM method. The experimental results show that IQPSO can obtain better optimization effect, and IQPSO-SVM has good generalization performance.

Massive Spatial Data Parallel Processing Technology for Vector Tiles

LI Han1, HU Ming-xiao2, GONG Zhi-hong3, FAN Bing-jun3

2018, 0(09): 32. doi:10.3969/j.issn.1006-2475.2018.09.007

Asbtract ( 193 )

References | Related Articles | Metrics

With the development of map mapping technology, the amount of original map data shows an explosive growth, and there is a higher demand for the spatial and temporal efficiency of map data processing. Aiming at this demand, this paper proposes a two-level parallel processing technique for vector tiles that include vector data cut-plan parallelism and vector tile upload parallelism. The experimental results show that this processing technique has a significant effect on improving the efficiency of vector data cropping and vector tile uploading. In the efficient processing of massive vector data, the proposed method is feasible.

Evaluation Method for Software Testing Quality

LI Jun-feng, GU Bin-bing, LI Hai-hao

2018, 0(09): 38. doi:10.3969/j.issn.1006-2475.2018.09.008

Asbtract ( 259 )

References | Related Articles | Metrics

The research status of current software testing quality evaluation and some misunderstandings are analyzed. Combining with the work experiences, based on easy to measure, effective and feasible, relatively impartial, we put forward a testing quality evaluation method according to the testing documents, testing adequacy, the results of spot test and testing efficiency. Finally, the software testing quality evaluation is validated through an actual project, and the next research direction of this method is summarized.

Vessel Trajectory Outlier Detection Algorithm Based on Adaptive Threshold

HAN Zhao-rong1,2,3, XU Guang-luan2,3, HUANG Ting-lei2,3, REN Wen-juan2,3

2018, 0(09): 42. doi:10.3969/j.issn.1006-2475.2018.09.009

Asbtract ( 284 )

References | Related Articles | Metrics

The rapid development of positioning technology has given rise to trajectory big data, and there are always obviously aberrant outliers in trajectories. Detecting outliers in trajectory data is crucial for improving data quality and the accuracy of subsequent knowledge discovery. The existing trajectory outlier detection algorithm is mainly the constant speed threshold approach, the method does not consider the change of the motion state at different moments of the object, can only detect a part of outliers whose velocity exceeds the specified threshold, and even leads to the detection error, the robustness of the algorithm is poor. Aiming at above problems, this paper proposes a Trajectory Outlier Detection Algorithm based on adaptive Threshold (TODAT). Taking full account of the motion information and observation noise impact of the object in a period of time, the TODAT algorithm applies the local threshold window and the mean filtering window to calculate the threshold and velocity, and also adds the economic speed threshold and the continuous outliers re-judgment mechanism. The experimental results based on real vessel data show that the proposed algorithm can get the adaptive threshold according to trajectory data, effectively detect all the outliers and greatly improve the quality of the trajectory data.

Uncategorized Text Verification Code Recognition Based on CTC Model

DU Wei, ZHUO Wu-neng

2018, 0(09): 48. doi:10.3969/j.issn.1006-2475.2018.09.010

Asbtract ( 225 )

References | Related Articles | Metrics

The security of verification code is an important part of securing the network. This paper uses deep learning to propose the long short-term memory (LSTM) network and connectionist temporal classification (CTC) model for intelligently recognizing the mainstream verification code images. The open source CAPTCHA verification code library is used to generate data sets, to simplify the verification code recognition model, to unify the speech recognition and text recognition methods, and to achieve the end-to-end model recognition. The proposed method has better performance under the condition of smaller training set.

Flower Image Classification Based on Convolutional Neural Network

ZHANG Xiao-feng, LIU Hong-zheng

2018, 0(09): 52. doi:10.3969/j.issn.1006-2475.2018.09.011

Asbtract ( 144 )

References | Related Articles | Metrics

Aiming at the problem of parameter redundancy and destruction of spatial structure information produced by fully connected layer in flower image classification using convolution neural network, this paper proposes an effective improvement method. Firstly, the n×n convolutional filters are replaced by 1×n and n×1 convolutional filters, then they are connected to the spatial pyramid pooling behind convolution layer to reduce the dimension and extract features, finally the probabilistic distribution is exported in Softmax classifier. Experimental results show that this method not only improves the accuracy, but also reduces the training time by half, which greatly improves the training speed.

Super-resolution Image Reconstruction Based on Adaptive Fractional Order Total Variation Regularization

LIU Ya-nan, PENG Ren-yong, WANG Lin

2018, 0(09): 56. doi:10.3969/j.issn.1006-2475.2018.09.012

Asbtract ( 202 )

References | Related Articles | Metrics

Super-resolution image reconstruction has important application value in various fields and has broad application prospects. It is an ill-posed problem to reconstruct the high-resolution image from the low-resolution image. The most effective method is to add the regularization term to solve it. This paper adds the fractional order total variation (FOTV) as the regularization term constraint solution space based on the traditional total variation (TV), and uses the texture detection function to determine the local features at different locations in the image, and selects the adaptive order. The alternating direction multiplier algorithm is used to divide the optimization function into multiple sub-problems and reduce the complexity of the operation. In this paper, the bi-regularization constraints of TV and the adaptive FOTV are used to adaptively reconstruct the texture detail information while removing the noise and sharpening edge. Experimental results show that compared with other methods, the proposed method improves the quality of image reconstruction, and both the PSNR and SSIM values are improved.

Monocular Visual Odometry Based on Improved BRISK Algorithm

FENG Jun, HUANG Duo-hui

2018, 0(09): 62. doi:10.3969/j.issn.1006-2475.2018.09.013

Asbtract ( 133 )

References | Related Articles | Metrics

In the traditional BRISK algorithm, a custom sampling pattern is used to describe the detected feature points, and a method based on the Hamming distance is used for feature matching. This feature point description and matching method of BRISK makes the low matching accuracy. Therefore, this paper proposes to combine SURF algorithm with high accuracy of matching and BRISK algorithm, and to use SURF descriptor and Euclidean distance-based matching method in BRISK feature point description and matching stage. The experimental results show that the accuracy of feature point matching is greatly improved when the time consumption of the algorithm is not greatly reduced. At the same time, the experiment also shows that the algorithm has good robustness.

Improved YOLO v2 for Target Recognition of Armored Vehicles

WANG Shu-guang, LYU Pan-fei

2018, 0(09): 68. doi:10.3969/j.issn.1006-2475.2018.09.014

Asbtract ( 181 )

References | Related Articles | Metrics

The technology of military target recognition is an important part of military information processing, which plays an important role in realizing the informatization and intelligentization of military equipment. In recent years, with the wide application of convolutional neural network in image recognition field, a variety of network structures based on image recognition task emerge in an endless stream. So it is of great practical significance and military application value to apply the new technology in military target recognition. Based on the YOLO v2 network which has the best recognition effect at present, this paper redefines the optimal number of anchors and their width and height dimensions by dimension clustering, and makes the armored vehicle data set with obvious features as the target area, so that the network can recognize the armored targets more accurately. Experimental results show that the method can effectively identify the specific armored targets in real time.

Discovering Domain Experts in Online Q&A Communities

LI Ke-lin

2018, 0(09): 72. doi:10.3969/j.issn.1006-2475.2018.09.015

Asbtract ( 137 )

References | Related Articles | Metrics

In this paper, we propose a question-answerer-topic model according to the generation procedure of answers in online Q&A communities, to model the topic distribution of question-answerer pairs. Then we calculate the professional level of users under different topics by incorporating the voting number of each answer in each question. Finally we propose an improved topic-sensitive PageRank model based on the thinking of link analysis to calculate the final expert score of each user under specific topics. The experiments and comparative analysis based on a real data set from a Chinese online Q&A community Zhihu in the field of artificial intelligence are carried out. The experimental results show that the proposed method obviously outperforms other existing expert finding methods.

Opinion Leader Mining Algorithms on Sina Weibo

LIU Jun-jie1, MA Chang2, SHAO Wei-long2, HAN Dong-hong2, XIA Li2

2018, 0(09): 80. doi:10.3969/j.issn.1006-2475.2018.09.016

Asbtract ( 197 )

References | Related Articles | Metrics

The current influence analysis algorithms are mostly based on network topology structure or user interaction information. However, a single method will lead to a large deviation in mining results. At present, there is no comprehensive and accurate influence mining method. Therefore, by extending the traditional PageRank algorithm, a new TCRank algorithm based on user interaction connection attribute is proposed for Sina Weibo. Secondly, three kinds of micro-blog opinion leader characteristics are designed, and their weighted summation is used to refine the candidate set of opinion leaders. At the same time, an opinion leader extraction algorithm based on emotional support of convolution neural network model is proposed to rank the candidate set of opinion leaders. Finally, the effectiveness of the proposed algorithm is verified by experiments.

Sectional Optimization Routing Algorithm Based on Geographic Location in UAV Network

FU Wei, ZHOU Xin-li

2018, 0(09): 87. doi:10.3969/j.issn.1006-2475.2018.09.017

Asbtract ( 151 )

References | Related Articles | Metrics

Due to the strong mobility of nodes, the possibility and randomness of network holes are increasing in UAV network. The general geographic assistant routing protocols often fail to satisfy the demand of the network. A sectional optimization mechanism based on packet feedback is proposed. This mechanism sets segment points based on the route forwarding strategy, optimizes the detour path part, and guarantees to respond to the existing network holes or optimize a new network hole duly. Considering the actual situation of the UAV cruise capability, comprehensive criteria is provided for selection of nodes according to the actual application requirements. The simulation shows that compared with the GPSR algorithm and BOPF algorithm, the proposed algorithm can optimize the transmission path, reduce end-to-end delay and improve network performance.

Replica Placement Algorithm in Sea-Cloud Collaboration Media Service System

BAO Sha-ru-la1,2, SUN Peng1,2, HAN Rui1,2, GUO Zhi-chuan1,2

2018, 0(09): 93. doi:10.3969/j.issn.1006-2475.2018.09.018

Asbtract ( 109 )

References | Related Articles | Metrics

For the problem of replica placement in Sea-Cloud collaboration media service system, this paper proposes a replica placement based on time matching algorithm (RPTM), which is based on the adjacent node collaborative distribution mechanism and aims at optimizing the transmission cost between nodes. The algorithm introduces the time matching factor in the heuristic greedy algorithm, to reduce the impact caused by node dynamic characteristics of sea nodes. The simulation results show that compared with the existing algorithms, the cost of transmission between the adjacent sea nodes is reduced by 10% to 31%.

Big-data Analysis of Network Search Frequency Based on Grey System Theory

LI Bin1, WU Qing-tao2

2018, 0(09): 98. doi:10.3969/j.issn.1006-2475.2018.09.019

Asbtract ( 132 )

References | Related Articles | Metrics

Big-data analysis is a process of applying descriptive, diagnostic, predictive, and prescriptive models for data to answer specific questions or to find new insights. Taking Baidu search index as a platform, and “Sanmenxia Polytechnic” as search keywords, this paper uses Web crawler software to intercept the weekly searching number of Baidu hot words from 2012 to 2017. Through the grey prediction model, the weekly searching frequency prediction equation is obtained. Compared with the predicted value of an element linear regression model, it’s verified that the prediction equation is reasonable and effective, and the number of keywords searched in the next two-year is predicted. Finally, through data chart analysis, “Sanmenxia Polytechnic” as Baidu search keywords has significant time characteristics: firstly, the total number of search is increasing every year, secondly, the peak and valley value every week in a year has an obvious fluctuation law. Combined with Baidu search index platform to analyze the periodic search distribution of keywords, the corresponding countermeasures are put forward.

Data Acquisition System for Agricultural Meteorological IOT

JU Shu-cun1,2, CHENG Wen-jie1,2, XU Jian-peng1,2, ZHOU Lu-yang1,2,

2018, 0(09): 105. doi:10.3969/j.issn.1006-2475.2018.09.020

Asbtract ( 150 )

References | Related Articles | Metrics

A data acquisition system for agricultural meteorological IOT based on STM32 processor is proposed, including microprocessor, power-supply module, sensors, GPS&GPRS, Flash memory and LCD touch screen. Experiments and demonstrations are conducted in 56 cities and counties of Anhui province. The results show that the monitoring system is operating stably; the collected data are complete and there is no obvious packet loss; the trend of data changing is consistent with the local agro-meteorological information, and it has a small deviation to the test results that using laboratory-standard instruments. The system can accurately monitor the agricultural meteorological conditions at the agricultural production site.

#br# Optimized Extended Kalman Filter Based on Learning Kernel Partial Least Squares

BAI Xiao-bo1, SHAO Jing-feng1, HE Zheng1, TIAN Jian-gang2

2018, 0(09): 110. doi:10.3969/j.issn.1006-2475.2018.09.021

Asbtract ( 163 )

References | Related Articles | Metrics

To solve the problem that the error of EKF estimation is larger with time sequence caused by uncertain parameters and inaccuracy noise information of none-liner system, the character of kernel partial least square which is independent of system equation parameters and noise information is used to optimize EKF. Firstly, measurement and convergence estimation value are as studying samples to build prediction model. And then, prediction values of KPLS and EKF are fused to estimate system status. Mean while, if convergence criterion of status estimation is true, then estimation values are as studying samples, and the sliding window is used to update kernel matrix, which makes KPLS have the ability to be predicted with time sequence. Else if convergence criterion of status estimation is false, then the measurement covariance will be updated. Finally, convergence and performance of KPLS-EKF are analyzed. The experimental results show that the proposed problem can be effectively solved by KPLS-EKF.

Sinusoidal Signal Frequency Estimation Method Based on Delay Cross-correlation

YAN Long-ji, XIAO Wei, CHEN Peng

2018, 0(09): 118. doi:10.3969/j.issn.1006-2475.2018.09.022

Asbtract ( 148 )

References | Related Articles | Metrics

To improve the frequency estimation precision of the sampled signals with non-integer period, a frequency estimation method for sinusoidal signal based on delay cross-correlation is proposed. Firstly, the sampled signal and its equal length delay signal are truncated. Secondly, the two segment signals are calculated by cross-correlation, and the error correction signal with non error term is constructed. Then, the coarse frequency estimation of the error correction signal is obtained, and the reference signal is generated. Finally, the error function is constructed according to the Cauchy Schwarz inequality, and the fine frequency estimation value is obtained by minimizing the error function. The results of simulation experiments indicate that the proposed method effectively reduces the influence of the non-integer period sampling, and improves the accuracy of frequency estimation, which is better than the frequency estimation method based on autocorrelation under the same conditions. The root mean square errors of the frequency estimation value are closer to the Cramer-Rao lower bound.

A Multiscale Convolutional Neural Network for Forex Trading Using Joint Feature Learning

CHEN Xi-yuan, ZHU Jia

2018, 0(09): 122. doi:10.3969/j.issn.1006-2475.2018.09.023

Asbtract ( 151 )

References | Related Articles | Metrics

Convolutional neural networks (CNNs) have revolutionized the field of computer vision. In this paper, we explore a particular application of CNNs: using CNNs to predict movements of forex prices from a picture of a time series of past price fluctuations, with the ultimate goal of using them for forex trading in order to make a profit. The main contribution of this research is to set up a novel architecture that uses multiscale CNNs to handle various kinds of features with a joint feature learning mechanism. Experimental results show our approach is more feasible compared with the basic CNNs using only image feature and other traditional machine learning methods.

Table of Content