Started in January,1974(Monthly)
Supervised and Sponsored by Chongqing Southwest Information Co., Ltd.
ISSN 1002-137X
CN 50-1075/TP
Current Issue
Volume 50 Issue 3, 15 March 2023
Special Issue of Knowledge Engineering Enabled By Knowledge Graph: Theory, Technology and System
SS-GCN:Aspect-based Sentiment Analysis Model with Affective Enhancement and Syntactic Enhancement
LI Shuai, XU Bin, HAN Yike, LIAO Tongxin
Computer Science. 2023, 50 (3): 3-11.  doi:10.11896/jsjkx.220700238
Abstract PDF(3238KB) ( 530 )   
References | Related Articles | Metrics
Aspect-based sentiment analysis(ABSA),as a downstream application of knowledge graph,belongs to the fine-grained sentiment analysis task,which aims to understand the sentiment polarity of people on the evaluation target at the aspect level.Relevant research in recent years has made significant progress,but existing methods focus on exploiting sequentiality or syntactic dependency constraints within sentences,and do not fully exploit the type of dependencies between context words and aspect words.In addition,the existing graph-based convolutional neural network models have insufficient ability to retain node features.In response to this problem,firstly,based on the syntactic dependency tree,this paper fully excavates the dependency types between context words and aspect words,and integrates them into the construction of the dependency graph.Second,we define a “sensitive relation set”,which is used to construct auxiliary sentences to enhance the correlation between specific context words and aspect words,and at the same time,combined with the sentiment knowledge network SenticNet to enhance the sentence dependency graph,and then improve the construction of the graph neural network.Finally,a context retention mechanism is introduced to reduce the information loss of node features in the multilayer graph convolution neural network.The proposed SS-GCN model fuses the syntactic and contextual representations learned in parallel to accomplish sentiment enhancement and syntactic enhancement,and extensive experiments on three public datasets demonstrate the effectiveness of SS-GCN.
Knowledge Graph-to-Text Model Based on Dynamic Memory and Two-layer Reconstruction Reinforcement
MA Tinghuai, SUN Shengjie, RONG Huan, QIAN Minfeng
Computer Science. 2023, 50 (3): 12-22.  doi:10.11896/jsjkx.220700111
Abstract PDF(4618KB) ( 407 )   
References | Related Articles | Metrics
Knowledge Graph-to-Text is a new task in the field of knowledge graph,which aims to transform knowledge graph into readable text describing these knowledge.With the deepening of research in recent years,the generation technology of Graph-to-Text has been applied to the fields of product review generation,recommendation explanation generation,paper abstract generation and so on.The translation model in the existing methods adopts the method of first-plan-then-realization,which fails to dynamically adjust the planning according to the generated text and does not track the static content planning,resulting in incohe-rent semantics before and after the text.In order to improve the semantic coherence of generated text,a Graph-to-Text model based on dynamic memory and two-layer reconstruction enhancement is proposed in this paper.Through three stages of static content planning,dynamic content planning and two-layer reconstruction mechanism,this model makes up for the structural difference between knowledge graph and text,focusing on the content of each triple while generating text.Compared with exis-ting generation models,this model not only compensates for the structural differences between knowledge graphs and texts,but also improves the ability to locate key entities,resulting in stronger factual consistency and semantics in the generated texts.In this paper,experiments are conducted on the WebNLG dataset.The results show that,compared with the current exis-ting models in the task of Graph-to-Text,the proposed model generates more accurate content planning.The logic between the sentences of the generated text is more reasonable and the correlation is stronger.The proposed model outperforms existing methods on me-trics such as BLEU,METEOR,ROUGE,CHRF++,etc.
Context-aware Temporal Knowledge Graph Completion Based on Relation Constraints
WANG Jingbin, LAI Xiaolian, LIN Xinyu, YANG Xinyi
Computer Science. 2023, 50 (3): 23-33.  doi:10.11896/jsjkx.220400255
Abstract PDF(3229KB) ( 349 )   
References | Related Articles | Metrics
The existing temporal knowledge graph completion models only consider the structural information of the quadruple itself,ignoring the implicit neighbor information and the constraints of relationships on entities,which leads to the poor perfor-mance of the models on the temporal knowledge graph completion task.In addition,some datasets exhibit unbalanced distribution in time,which makes it difficult for model training to achieve a good balance.To address these problems,the paper proposes a context-aware model based on relation constraints(CARC).CARC solves the problem of an unbalanced distribution of datasets in time through an adaptive time granularity aggregation module and uses a neighbor-aggregator to integrate contextual information into entity embeddings to enhance the embedding representation of the entity.In addition,the quadruple relation constraint mo-dule is designed to make the embeddings of entities with the same relational constraints close to each other,while those with diffe-rent relational constraints are far away from each other,which further enhances the embedding representation of entities.Extensive experiments are conducted on several publicly available temporal datasets,and the experimental results prove the superiority of the proposed model.
Multi-information Optimized Entity Alignment Model Based on Graph Neural Network
CHEN Fuqiang, KOU Jiamin, SU Limin, LI Ke
Computer Science. 2023, 50 (3): 34-41.  doi:10.11896/jsjkx.220700242
Abstract PDF(2351KB) ( 328 )   
References | Related Articles | Metrics
Entity alignment is a key step in knowledge fusion,which aims to discover entity pairs with corresponding relations between knowledge graphs.Knowledge fusion enables a more extensive and accurate services for further knowledge graph applications.However,the entity names and relations are used insufficiently by most of the state-of-the-art models of entity alignment.After obtaining the vector representation of the entity,generally the alignment relations among the entities are obtained through single iterative strategy or direct calculation,while ignoring some valuable information,so that the result of entity alignment is not ideal.In view of the above problems,a multi-information optimized entity alignment model based on graph neural network(MOGNN) is proposed.Firstly,the input of the model fuses word information and character information in the entity name,and the vector representation of relations is learnt through attention mechanism.After transmitting the information by utilizing relations,MOGNN corrects the initial entity alignment matrix based on the pre-alignment results of entities and relations,and finally employs the deferred acceptance algorithm to further correct the misaligned results.The proposed model is validated on three subsets of DBP15K,and compared with the baseline models.Compared with the baseline models,Hits@1 increases by 4.47%,0.82% and 0.46%,Hits@10 and MRR have also achieved impressive results,and the effectiveness of the model is further verifies by ablation experiments.Therefore,more accurate entity alignment results can be obtained with the proposed model.
BGPNRE:A BERT-based Global Pointer Network for Named Entity-Relation Joint Extraction Method
DENG Liang, QI Panhu, LIU Zhenlong, LI Jingxin, TANG Jiqiang
Computer Science. 2023, 50 (3): 42-48.  doi:10.11896/jsjkx.220600239
Abstract PDF(2098KB) ( 302 )   
References | Related Articles | Metrics
Named entity-relation joint extraction refers to extracting entity-relation triples from unstructured text.It's an important task for information extraction and knowledge graph construction.This paper proposes a new method--BERT-based global pointer network for named entity-relation joint extraction(BGPNRE).Firstly,the potential relation prediction module is used to predict the relations contained in the text,filters out the impossible relations,and limits the predicted relation subset for entity recognition.Then a relation-specific global pointer-net is used to obtain the location of all subject and object entities.Finally,a global pointer network correspondence component is designed to align the subject and object position into named entity-relation triples.This method avoids error propagation frompipeline model,and also solves the the redundancy of relation prediction,entity overlapping,and poor generalization of span-based extraction.Extensive experiments show that our model achieves state-of-the-art performance on NYT and WebNLG public benchmarks with higher performance gain on multi relations and entities overlapping.
Study on Graph Neural Networks Social Recommendation Based on High-order and Temporal Features
YU Jian, ZHAO Mankun, GAO Jie, WANG Congyuan, LI Yarong, ZHANG Wenbin
Computer Science. 2023, 50 (3): 49-64.  doi:10.11896/jsjkx.220700108
Abstract PDF(4274KB) ( 256 )   
References | Related Articles | Metrics
Cross-item social recommendation is a method of integrating social relationships into the recommendation system.In social recommendation,user is the bridge connecting user-item interaction graph and user-user social graph.So user representation learning is essential to improve the performance of social recommendation.However,existing methods mainly use static attributes of users or items and explicit friend information in social networks for representations learning,and the temporal information of the interaction between users and items and their implicit friend information are not fully utilized.Therefore,in social recommendation,effective use of temporal information and social information has become one of the important research topics.This paper focuses on the temporal information of the interaction between users and items,and gives full play to the advantages of social network,modeling the user's implicit friends and item's social attributes.This paper proposes a novel graph neural networks social recommendation based on high-order and temporal features,referred to as HTGSR.Firstly,the framework uses gated recurrent unit to model item-based user representations to reflect the user's recent preferences,and defines a high-order mo-deling unit to extract the user's high-order connected features and obtain the user's implicit friend information.Secondly,HTGSR uses attention mechanism to obtain social-based user representation.Thirdly,the paper proposes different ways to construct item's social networks,and uses the attention mechanism to obtain item representations.Finally,the user's and item's representations are input to the MLP to complete the user's rating prediction for the item.The paper conducts specific experiments on two public and real-world datasets,and compares the experimental results with different recommendation algorithms.The results show that the HTGSR has achieved good results on the two datasets.
Method of Java Redundant Code Detection Based on Static Analysis and Knowledge Graph
LIU Xinwei, TAO Chuanqi
Computer Science. 2023, 50 (3): 65-71.  doi:10.11896/jsjkx.220700240
Abstract PDF(1429KB) ( 208 )   
References | Related Articles | Metrics
Redundant code is common in commercial and open source software,and its presence can increase memory footprint,affect code maintainability,and increase maintenance costs.Rapid type analysis algorithm is a common static analysis method in Java redundant code detection,but it still has some shortcomings in virtual method analysis.XTA is a call graph construction algorithm with high precision and efficiency in handling virtual method calls.A method based on XTA call graph construction algorithm is proposed to detect redundant code in Java code.This method is implemented in a prototype tool called redundant code Detection(RCD),and the knowledge graph is constructed to assist manual review to improve the efficiency of manual review and the reliability of redundant code detection.RCD is compared with three other redundant code detection tools by experiments on four open source Java applications.Experimental results show that RCD improves the accuracy of detecting redundant codes by 1%~30% compared with other tools,and improves the integrity of detecting redundant virtual methods by about 4%.
Fine-grained Semantic Knowledge Graph Enhanced Chinese OOV Word Embedding Learning
CHEN Shurui, LIANG Ziran, RAO Yanghui
Computer Science. 2023, 50 (3): 72-82.  doi:10.11896/jsjkx.220700249
Abstract PDF(2405KB) ( 294 )   
References | Related Articles | Metrics
With the expansion of the scope in informatization fields,lots of text corpora in specific fields continue to appear.Due to the impact of security and sensitivity,the text corpora in these specific fields(e.g.,medical records corpora and communication corpora) are often small-scaled.It is difficult for traditional word embedding learning methods to obtain high-quality embeddings on these corpora.On the other hand,there may exist many out-of-vocabulary words in these corpora when using the existing pre-training language models directly,for which,many words cannot be represented as vectors and the performance on downstream tasks are limited.Many researchers start to study how to infer the semantics of out-of-vocabulary words and obtain effective out-of-vocabulary word embeddings based on fine-grained semantic information.However,the current models utilizing fine-grained semantic information mainly focus on the English corpora and they only model the relationship among fine-grained semantic information by simple ways of concatenation or mapping,which leads to a poor model robustness.Aiming at addressing the above problems,this paper first proposes to construct a fine-grained knowledge graph by exploiting Chinese word formation rules,such as the characters contained in Chinese words,as well as the character components and pinyin of Chinese characters.The know-ledge graph not only captures the relationship between Chinese characters and Chinese words,but also represents the multiple and complex relationships between Pinyin and Chinese characters,components and Chinese characters,and other fine-grained semantic information.Next,the relational graph convolution operation is performed on the knowledge graph to model the deeper relationship between fine-grained semantics and word semantics.The method further mines the relationship between fine-grained semantics by the sub-graph readout,so as to effectively infer the semantic information of Chinese out-of-vocabulary words.Experimental results show that our model achieves better performance on specific corpora with a large proportion of out-of-vocabulary words when applying to tasks such as word analogy,word similarity,text classification,and named entity recognition.
Survey of Medical Knowledge Graph Research and Application
JIANG Chuanyu, HAN Xiangyu, YANG Wenrui, LYU Bohan, HUANG Xiaoou, XIE Xia, GU Yang
Computer Science. 2023, 50 (3): 83-93.  doi:10.11896/jsjkx.220700241
Abstract PDF(2148KB) ( 357 )   
References | Related Articles | Metrics
In the process of digitisation of medical data,choosing the right technology for efficient processing and accurate analysis of medical data is a common problem faced by the medical field today.The use of knowledge graph technology with the excellent association and reasoning capabilities to process and analyse medical data can better enable applications such as wise information technology of medicine and aided diagnoses.The complete process of constructing a medical knowledge graph includes know-ledge extraction,knowledge fusion and knowledge reasoning.Knowledge extraction can be subdivided into entity extraction,relationship extraction and attribute extraction,while knowledge fusion mainly includes entity alignment and entity disambiguation.Firstly,the constructiontechnologies and practical applications of medical knowledge graphs are summarised,and the development of the technologies is clarified for each specific construction process.On this basis,the relevant techniques are introduced,and their advantages and limitations are explained.Secondly,introducing several medical knowledge graphs that are being successfully applied.Finally,based on the current state of technology and applications of knowledge graphs in the medical field,future research directions for knowledge graphs in technology and applications are given.
Survey of Knowledge Graph Reasoning Based on Representation Learning
LI Zhifei, ZHAO Yue, ZHANG Yan
Computer Science. 2023, 50 (3): 94-113.  doi:10.11896/jsjkx.220900136
Abstract PDF(4422KB) ( 337 )   
References | Related Articles | Metrics
Knowledge graphs describe objective knowledge in the real world in a structured form,and are confronted with issues of completeness and newly-added knowledge.As an important means of complementing and updating knowledge graphs,know-ledge graph reasoning aims to infer new knowledge based on existing knowledge.In recent years,the research on knowledge graph reasoning based on repre-sentation learning has received extensive attention.The main idea is to convert the traditional reasoning process into semantic vector calculation based on the distributed representation of entities and relations.It has the advantages of fast calculation efficiency and high reasoning performance.In this paper,we review the knowledge graph reasoning based on repre sentation learning.Firstly,this paper summarizes the symbolic representation,data set,evaluation metric,training method,and evaluation task of knowledge graph reasoning.Secondly,it introduces the typical methods of knowledge graph reasoning,including translational distance and semantic matching methods.Thirdly,multi-source information fusion-based knowledge graph reasoning methods are classified.Then,neural network-based reasoning methods are introduced including convolutional neural network,graph neural network,recurrent neural network,and capsule network.Finally,this paper summarizes and forecasts the future research direction of knowledge graph reasoning.
Database & Big Data & Data Science
Cross-network User Identification Based on Multiple Spatio-Temporal Trajectory Features
LIU Hong, ZHU Yan, LI Chunping
Computer Science. 2023, 50 (3): 114-120.  doi:10.11896/jsjkx.211200287
Abstract PDF(2602KB) ( 2279 )   
References | Related Articles | Metrics
With the flourishing of location-based social networks,users’mobile behavior data has been greatly enriched,which promotes the research on user identification based on spatio-temporal data.User identification in cross-location social networks emphasizes learning the correlation between time and space sequences of different platforms,aiming at discovering the accounts registered by the same user on different platforms.In order to solve the problems of data sparsity,low quality and spatio-temporal mismatch faced by existing researches,a recognition algorithm UI-STDD combining bidirectional spatio-temporal dependence and spatio-temporal distribution is proposed.The algorithm mainly consists of three modules:the space-time sequence module is combined with the bidirectional long short-term memory network of paired attention to describe user movement patterns;the time preference module defines the user personalized mode from coarse and fine granularity;the spatial location module mines local and global information of location points to quantify spatial proximity.Based on the user trajectory pair features obtained by the above modules,a multi-layer feedforward network is used in UI-STDD to distinguish whether two accounts across the network corres-pond to the same person in real life.To verify the feasibility and effectiveness of UI-STDD,experiments are carried out on three publicly available datasets.Experimental results show that the proposed algorithm can improve the user identification rate based on spatio-temporal data,and the average F1 value is more than 10% higher than the optimal comparison method.
Study on Air Traffic Flow Recognition and Anomaly Detection Based on Deep Clustering
RAO Dan, SHI Hongwei
Computer Science. 2023, 50 (3): 121-128.  doi:10.11896/jsjkx.220100086
Abstract PDF(4271KB) ( 2309 )   
References | Related Articles | Metrics
Aiming at the problem that traditional clustering algorithms cannot capture the implicit relationship of high-dimen-sional trajectory data in low-dimensional space,and it is difficult to define appropriate similarity measures to consider both local and global features of trajectories,a multivariate trajectory deep clustering(MTDC) framework based on deep neural network(DNN) is proposed and used for air traffic flow recognition and anomaly detection.The framework mainly includes an asymmetric autoencoder and a custom trajectory clustering layer.The autoencoder is mainly composed of 1D convolutional neural network and bi-directional long short-term memory to learn the feature representation of the original input in the low-dimensional latent space.The trajectory clustering layer realizes clustering by calculating the Q distribution of samples in the hidden space.Combined with reconstruction loss of autoencoder and trajectory clustering Q distribution,a new anomaly score is defined for anomaly trajectory detection.The results of experiments using real trajectory data based on automatic dependent surveillance-broadcast(ADS-B) show that the proposed framework is effective for air traffic flow recognition and can detect anomaly trajectories that are mea-ningful and interpretable.
Attention-aware Multi-channel Graph Convolutional Rating Prediction Model for Heterogeneous Information Networks
ZHOU Mingqiang, DAI Kailang, WU Quanwang, ZHU Qingsheng
Computer Science. 2023, 50 (3): 129-138.  doi:10.11896/jsjkx.220300004
Abstract PDF(4257KB) ( 2234 )   
References | Related Articles | Metrics
Heterogeneous information network(HIN) contains rich semantic information,and the use of HIN for rating prediction has become an important way to alleviate the problem of data sparsity in recommender systems.However,the traditional methods using meta-paths to extract HIN semantic information ignore the rating information on the edges,making the meta-paths unable to accurately capture the semantic similarity between users and recommended items.And these methods also fail to distinguish the importance of different meta-paths.To address the two problems,rating constrained meta-path is proposed to obtain more accurate HIN semantic information which is then used to construct multi-layer homogeneous networks for users and items.Then,a neural network for rating prediction is designed by combining graph convolutional network and attention mechanism,which effectively represents various semantic information in HIN through multi-channel graph convolutional networks and distinguishes the importance of different meta-paths by using an attentional fusion function.Furthermore,the proposed model also integrates the attribute information of users and items to improve the accuracy of rating prediction.Experimental results on Douban Book and Yelp datasets show that the proposed model is significantly better than the comparative baseline models,especially in the case of sparse data,and the root mean square error reduces by up to 50% compared to the baseline model,thus verifying the superiority of the proposed model.
ECG Abnormality Detection Based on Label Co-occurrence and Feature Local Pertinence
HAN Jingyu, QIAN Long, GE Kang, MAO Yi
Computer Science. 2023, 50 (3): 139-146.  doi:10.11896/jsjkx.220200004
Abstract PDF(1927KB) ( 2237 )   
References | Related Articles | Metrics
Automatic electrocardiogram(ECG) abnormality detection is a multi-label classification problem,which is commonly solved by training a binary-relevance classifier for each abnormality.Due to the large number of abnormalities,the complex correlations between features and abnormalities,and those among different abnormalities,existing methods’ performance is not satis-fying.To make full use of the dependencies between features and abnormalities,this paper proposes a novel abnormality detection method based on label co-occurrence and feature local pertinence(LCFP).Firstly,we set up a consolidated feature space consisting of both the macro-features and micro-features based on the label co-occurrence and features’ pertineance.The macro-features are constructed with a clustering approach based on Dirichlet process mixture model(DPMM),thus distinguishing differentco-occurrence label sets.The micro-features are a subset of primitive features,which serves to distinguish between the labels in the same labelset.Next,we train a one-versus-all classifier which returns a relevance probability.Secondly,to make use of the diffe-rent correlation degrees among different abnormalities,we propose to differ the relevant labels from the irrelevant ones based on the sorting according to the probabilities given by the classifiers.In particular,we propose to exploit the Beta distribution to adaptively learn the anchor thresholds and correlation thresholds,thus determining the relevant labels of an instance.Our LCFP me-thod is a universal way to detect every possible ECG abnormalities,which effectively improves the detection accuracy.The results on two real datasets show that our method can achieves an improvement of 4% and 22.4%,respectively,in terms of F1,which proves the effectiveness of our method.
Mining Negative Sequential Patterns with Periodic Gap Constraints
WANG Zhulin, WU Youxi, WANG Yuehua, LIU Jingyu
Computer Science. 2023, 50 (3): 147-154.  doi:10.11896/jsjkx.211200248
Abstract PDF(1980KB) ( 2194 )   
References | Related Articles | Metrics
Sequential pattern mining with gap constraints is a special form of sequential pattern mining,which can reveal frequent subsequences in a certain gap.However,the current sequential pattern mining methods with gap constraints only focus on positive sequential pattern mining,and ignore the missing behavior in a series of events.To solve this problem,a negative sequential pattern method with periodic gap constraints(NSPG) mining is explored,which can reflect the relationship between elements more flexibly.To solve the problem of NSPG mining,this paper proposes an NSPG-INtree(incomplete nettrees) algorithm,which includes two key steps:candidate pattern generation and support calculation.For candidate pattern generation,to reduce the number of candidate patterns,the algorithm uses a pattern join strategy.For support calculation,to improve the efficiency and reduce space consumption,the algorithm employs an incomplete nettree structure to calculate the supports of patterns.Experimental results show that NSPG-INtree not only has high mining efficiency,but also can mine positive and negative sequential patterns with gap constraints.NSPG-INtree can find 209% ~ 352% more patterns than other gap-constrained sequential pattern mining algorithms.Moreover,NSPG-INtree can reduce the running time by 6%~38% than other competitive algorithms with different stra-tegies.
Nodes’ Ranking Model Based on Influence Prediction
DUAN Shunran, YIN Meijuan, LIU Fenlin, JIAO Longlong, YU Lanlan
Computer Science. 2023, 50 (3): 155-163.  doi:10.11896/jsjkx.211200261
Abstract PDF(3807KB) ( 2181 )   
References | Related Articles | Metrics
The ranking of nodes’ influence has always been a hot issue in the research area of complex networks.Susceptible-infected-recovered(SIR) model is an ideal nodes’ influence ranking method,which is commonly used to evaluate other nodes’ in-fluence ranking methods.But it is difficult to be applied in practice due to its high time complexity.This paper proposes a nodes’ influence ranking model based on sir value learning.Both the local structure and global structure information of nodes are used as features in the model.The sir value learning model is constructed by means of a deep learning model,which is trained on nodes’ features and sir data set in synthetic graphs with the same size.The trained model can predict sir value based on nodes’ features,and then rank nodes’ influence based on predicted sir.In this paper,a specific nodes’ influence ranking method is implemented based on the proposed model,and experiments are carried out on five real networks to verify the effectiveness of the method.The results show that the accuracy and monotonicity of nodes’ influence ranking results are improved compared with degree centrality,Kshell and Weighted Kshell degree neighborhood.
Improving RNA Base Interactions Prediction Based on Transfer Learning and Multi-view Feature Fusion
WANG Xiaofei, FAN Xueqiang, LI Zhangwei
Computer Science. 2023, 50 (3): 164-172.  doi:10.11896/jsjkx.211200186
Abstract PDF(5749KB) ( 2241 )   
References | Related Articles | Metrics
RNA base interactions play an important role in maintaining the stability of its three-dimensional structure,and accurate prediction of base interactions can help predict the three-dimensional structure of RNA.However,due to the small amount of data,the model could not effectively learn the feature distribution of the training data,and existing data characteristics(symmetry and class imbalance) affect the performance of the RNA base interactions prediction model.Aiming at the problems of insufficient model learning and data characteristics,a high-performance RNA base interactions prediction method called tpRNA is proposed based on deep learning.tpRNA introduces transfer learning in RNA base interactions prediction task to weak the influence of insufficient learning in the training process due to the small amount of data,and an efficient loss function and feature extraction module is proposed to give full play to the advantages of transfer learning and convolutional neural network in feature learning to alleviate the problem of data characteristics.Results show that transfer learning can reduce the model deviation caused by less data,the proposed loss function can optimize the model training,and the feature extraction module can extract more effective features.Compared with the state-of-the-art method,tpRNA also has significant advantages in the case of low-quality input features.
Graph Attention Deep Knowledge Tracing Model Integrated with IRT
DONG Yongfeng, HUANG Gang, XUE Wanruo, LI Linhao
Computer Science. 2023, 50 (3): 173-180.  doi:10.11896/jsjkx.211200134
Abstract PDF(2077KB) ( 2217 )   
References | Related Articles | Metrics
Knowledge tracing aims to trace students’ knowledge state(the degree of knowledge) based on their historical answer performance in real time and predict their future answer performance.The current research only explores the direct influence of the question or concept itself on the performance of students’ answering questions,while often ignores the indirect influence of the deep-level information in the questions and the concepts contained on the performance of students’ answering questions.In order to make better use of these deep-level information,a graph attention deep knowledge tracing model integrated with IRT(GAKT-IRT) is proposed,which integrates item response theory(IRT).The graph attention network is applied to the field of knowledge tracing and uses IRT to increase the interpretability of the model.First,obtain the deep-level feature representation of the problem through the graph attention network layer.Next,model students’ knowledge state based on their historical answer sequence that combines the in-depth information.Then,use IRT to predict students’ future answer performance.Results of comparative experiments on 6 open real online education datasets prove that the GAKT-IRT model can better complete the knowledge tracing task and has obvious advantages in predicting the future performance of students in answering questions.
Computer Graphics & Multimedia
Shooting Event Detection of Free Kick in Soccer Video Based on Rule Reasoning
HUA Xiaofeng, FENG Na, YU Junqing, HE Yunfeng
Computer Science. 2023, 50 (3): 181-190.  doi:10.11896/jsjkx.220300062
Abstract PDF(2971KB) ( 150 )   
References | Related Articles | Metrics
Soccer video event detection is of great significance to video retrieval.However,there are fewer events in soccer videos,and these events mainly occur in the far-view shot,which makes it difficult to capture key players and key actions,making soccer event detection more difficult.In recent years,methods based on deep learning have made some progress in soccer video event detection,but the learning ability of the high-level semantic of the event is still insufficient and the detection results need to be further improved.Therefore,how to improve the accuracy of soccer video event detection is an urgent problem to be solved.Taking the shooting event of free kick(free-kick shot event) as the research object,an event detection model combining soccer rules and deep learning is proposed.To have a deeper understanding of the inherent characteristics of the free-kick shot event,the event rules are manually summarized and verified on the public soccer dataset,and the corresponding application scenarios are also proposed.For the problem of too few events in soccer videos,rule-based initial localization algorithm is proposed to preprocess the videos.Through the combination and application of multiple rules,the location where the free-kick shot event may occur is initially located from the original video,which is used as the input of the deep learning model for further prediction.The proposed mo-del is compared with other models on the public soccer dataset.Experimental results show that the proposed model achieves the best results,with the accuracy rate of 78% and the recall rate of 81.25%.Compared with other models,the improvement in accuracy is particularly prominent.It can be seen that the free-kick shot event detection model that combines soccer rules and deep learning effectively improves the performance of free-kick shot event detection and provides a basic reference for further research on the detection of other events in soccer videos.
Sound Event Joint Estimation Method Based on Three-dimension Convolution
MEI Pengcheng, YANG Jibin, ZHANG Qiang, HUANG Xiang
Computer Science. 2023, 50 (3): 191-198.  doi:10.11896/jsjkx.220500259
Abstract PDF(3439KB) ( 130 )   
References | Related Articles | Metrics
Sound event localization and detection(SELD) is widely used in monitoring and anomaly detection tasks.Deep learning methods represented by convolutional recurrent neural networks(CRNN) can be realized to improve the performance of SELD.In order to improve the system localization and detection performance,a method based on 3D Convolution feature extraction,called SELD3Dnet,is proposed.The amplitude and phase spectra of input multi-channel acoustic signal are calculated,and the deep feature representation is extracted by multiple 3D Convolution modules.Recurrent neural networks and the fully connected layers are adopted to estimate the type of sound events and their localization.In processing multi-channel acoustic signals,three-dimensional(3D) convolution can carry out convolution calculation of time,frequency and signal channel simultaneously,so that the correlation between signal channels can be exploited to the maximum extent.Comparative experiments are conducted on TUT2018 dataset and TAU2019 dataset,and the results show that the comprehensive performance of the proposed method is significantly improved on TUT2018 REAL and TAU2019 MREAL datasets.The F1 index of acoustic event detection on TUT2018 REAL dataset significan-tly improves by 13.9% and frame accuracy by 21.1%,while the F1 index on TAU2019 MREAL dataset significantly improves by 10.8% and frame accuracy by 14.4%.It is verified that the proposed method can effectively overcome the influence of reverberation existing in real-life scenes.
Segmentation Method of Edge-guided Breast Ultrasound Images Based on Feature Fusion
BAI Xuefei, MA Yanan, WANG Wenjian
Computer Science. 2023, 50 (3): 199-207.  doi:10.11896/jsjkx.211200294
Abstract PDF(3397KB) ( 184 )   
References | Related Articles | Metrics
Due to the problems of blurred edges,excessive speckle noise,and low contrast in breast ultrasound images,an edge-guided multi-scale selective kernel U-Net(EMSK U-Net) method that fuses multiple features is proposed.The U-Net network has a symmetrical encoder-decoder structure,which can achieve better segmentation results on medical images with a small amount of data.EMSK U-Net adopts a network structure based on it,which combines dilated convolution with traditional convolution to form a selective kernel module,and applies it to the encoding path of the symmetric structure.Meanwhile,in the encoding part,EMSK U-Net performs the task of edge detection by extracting selective kernel features during down sampling.Through these methods,the spatial information of the image is enriched and the edge information of the image is refined,which effectively alleviates the difficult problem of segmentation caused by speckle noise and edge blur in breast ultrasound images,and the detection accuracy of small targets will also be improved to a certain extent.After that,in the decoding path of U-Net,EMSK U-Net obtains rich deep semantic information by building a multi-scale feature weighted aggregation module,realizes more information interaction between deep and shallow layers,and reduces the problem of low contrast.In general,EMSK U-Net jointly guides the segmentation of the network by complementing various information such as encoding part of the spatial information,edge information and decoding part of the depth feature information,so that the segmentation performance has been well improved.Experiments are conducted on three public breast ultrasound image datasets,and the results show that compared with other classical medical image segmentation methods and breast ultrasound segmentation methods,the EMSK U-Net algorithm performs well in various indicators.The performance of breast ultrasound image segmentation task has been significantly improved.
Polarized Self-attention Constrains Color Overflow in Automatic Coloring of Image
LIU Hang, PU Yuanyuan, LYU Dahua, ZHAO Zhengpeng, XU Dan, QIAN Wenhua
Computer Science. 2023, 50 (3): 208-215.  doi:10.11896/jsjkx.220100149
Abstract PDF(5154KB) ( 158 )   
References | Related Articles | Metrics
Auto coloring transforms grayscale images into reasonable colored versions of natural color images,allowing the restoration of color for old photographs,black-and-white films,videos,etc.Therefore,it is widely concerned in the realms of computer vision and graphics.Nevertheless,allocating colors to grayscale images is a highly challenging mission with a color overflow pro-blem.To address the problem,a technique for automatic coloring of images with polarized self-attention constrained color overflow is proposed.At first,separating instances in the foreground from the background minimizes the coloring effect of the background against the foreground,to mitigate the color overflow among the foreground and background.Second,the polarized self-attention block splits the features into color channels and spatial locations for more accurate and specific coloring,which reduces the color overflow within the global image,instance objects.At last,the fusion module is combined to integrate the global features and instance features through different weights to accomplish the ultimate coloring.Experiment results show that the main indexes FID and LPIPS are improved by 9.7% and 10.9% respectively,and the indexes SSIM and PSNR are optimal compared with ChromaGAN and MemoGAN.
Leaf Classification and Ranking Method Based on Multi-granularity Feature Fusion
LIU Songyue, WANG Huan
Computer Science. 2023, 50 (3): 216-222.  doi:10.11896/jsjkx.211100203
Abstract PDF(2674KB) ( 147 )   
References | Related Articles | Metrics
Much work has long been devoted to plant leaf classification,but these methods cannot be applied well in real applications,though they may achieve good results in public datasets.Moreover,they are hardly employed to more complex problems,e.g.leaf ranking,which requires the classification of leaves first and then ranking leaves of the same class.This paper proposes a new model for plant leaf classification as well as leaf ranking,which focuses on the granularity information of leaves and integrates multi-level granularity from coarse to fine.Specifically,the model contains two branches,coarse-grained and fine-grained,which are linked by a coarse-fine hybrid loss,prompting the model to progressively learn a coarse-to-fine representation.A multi-step training approach is used,with different levels of features extracted at each step,therefore enabling the fusion of shallow features with deep features.In addition,a geometric channel attention module,which consists of a spatial transformation and a bili-near attention pooling module,is proposed to allow our model to focus on more discriminative local regions in the image and extract more discriminative features.Our method achieves 99.8% and 99.7% classification accuracy on two publicly available leaf classification datasets,Flavia leaf and Swedish leaf,respectively,and 71.9% classification accuracy on our constructed tobacco leaf ranking dataset,both outperform the state-of-the-art methods.
Classification of Oil Painting Art Style Based on Multi-feature Fusion
XIE Qinqin, HE Lang, XU Ruli
Computer Science. 2023, 50 (3): 223-230.  doi:10.11896/jsjkx.211200110
Abstract PDF(2615KB) ( 137 )   
References | Related Articles | Metrics
The existing oil painting art style classification algorithms ignore the influence of the main area and the overall effect on the art style.Aiming at this problem,this paper proposes a new oil painting classification algorithm based on multi-feature fusion classifier(MFFC).Firstly,based on the common arrangement form of oil painting art elements,this paper designs the overlapping image block method.This method extracts spatial features of oil paintings to make up for the lack of composition style in existing algorithms.And it also can be used to distinguish the subject area from the background area.Secondly,the spatial features and the underlying features are combined in series to increase the location information of the elements in the picture.Finally,the spatial voting method is designed to give priority to the classification result of the main area as the output result of the algorithm.This is to highlight the role of oil painting subject area in the classification and realize the automatic classification of oil painting art style.Tested on the data set created by the FS-classifier model,its accuracy,precision,recall,F1-score and AUC reaches 96.92%,63.69%,98.75%,98.57% and 0.917,respectively.Compared with FS-classifier,the result increases by 6.72%,5.85%,9.05%,7.1% and 0.128,respectively.When tested on WIKIART and compared with other six algorithms,the accuracy improves by 13.27%,at least.The results show that the proposed algorithm can effectively improve the performance of spatial features for oil painting art style classification task,and has good practical value.
SSD Object Detection Algorithm with Cross-layer Fusion and Receptive Field Amplification
ZHANG Weiliang, CHEN Xiuhong
Computer Science. 2023, 50 (3): 231-237.  doi:10.11896/jsjkx.211100281
Abstract PDF(3432KB) ( 165 )   
References | Related Articles | Metrics
In view of the lack of information interaction between different layers of single shot multibox detector(SSD) and the limitation of the model's receptive field,an improved SSD object detection algorithm,named ESSD(enhanced SSD),is proposed to improve the accuracy of object detection.First of all,using the original multi-scale feature map in the SSD model and using the idea of feature pyramid networks(FPN),a cross-layer information interaction module is designed,which enhances the semantic information capabilities of different layers and reduces the information difference of different layers.Then,in order to improve the receptive field and multi-scale detection capabilities of the model,a receptive field amplification module is designed.Finally,the batch normalization layer is used to reduce the training time and improve the convergence speed of the model.In order to evaluate the effectiveness of ESSD,experiments are conducted on the PASCAL VOC2007 and PASCAL VOC2012 test sets.Experimental results show that on the PASCAL VOC2007 data set,its mAP is 82.1% and the detection speed is 15.7FPS.Compared with the original SSD512,its mAP increases by 2.3%;on the PASCAL VOC2012 test set,its mAP reaches 80.6%,which is also 2.1% higher than SSD512.Experiments have proved that the ESSD detector can still meet the real-time performance under the condition of high detection accuracy.
Study on Visual Dashboard Generation Technology Based on Deep Learning
CHEN Liang, WANG Lu, LI Shengchun, LIU Changhong
Computer Science. 2023, 50 (3): 238-245.  doi:10.11896/jsjkx.230100064
Abstract PDF(3758KB) ( 150 )   
References | Related Articles | Metrics
Dashboard is an important tool to support manufacturing enterprises in data visualization analysis and business decision-making.In order to solve the problems of users’ strong dependence on professional knowledge and complicated process iteration in the design and implementation of visual dashboards,a method for automatic recognition and generation of visual dashboards based on the YOLOv5s algorithm in deep learning technology is proposed.Firstly,based on the YOLOv5s algorithm,the visual chart components contained in the dashboard images and hand-drawn sketches are detected,and in order to address the problems of interference and false detection caused by irregular lines in hand-drawn sketches during the detection process,the CA attention mechanism is introduced to enhance the ability of the model to focus on important features and accurately locate the target,so as to improve the recognition accuracy of the model.Secondly,deploy the chart detection model in the web,and the server calls the encapsulated visual chart component code according to the model detection results to generate the initial dashboard of multi-component combination.Finally,a data visualization dashboard building platform is developed based on this web design,which provides users with detailed options to modify and configure the dashboard style and data,so that users can quickly build a complete dashboard.Through the collection of dashboard images generated by visualization tools such as Tableau and Power BI and hand-drawn dashboard sketches during the enterprise application process to form a dataset for experimental validation,the improved model increases the recognition accuracy by 2.1% compared to the original YOLOv5s model,and the mAP is 98.4%.The system deployment application verifies that the chart recognition method and the developed platform can effectively identify and generate the corresponding chart components to meet the basic needs of users.
Crowd Counting Network Based on Feature Enhancement Loss and Foreground Attention
Computer Science. 2023, 50 (3): 246-253.  doi:10.11896/jsjkx.220100219
Abstract PDF(6985KB) ( 157 )   
References | Related Articles | Metrics
Crowd counting aims to estimate the total number of people in an image and present its distribution accurately.The images in the relevant datasets usually involve a variety of scenes and include multiple people.To save labor,most datasets usually annotated each human head by a single point.However,the point labels cannot cover the full human head,which makes it difficult to converge the matching between the crowd feature and the distribution label,and the predicted values cannot be gathered in the foreground region,which seriously affects the density estimation map quality and count accuracy.To solve this problem,count loss is used to constrain the range of predictions on the full map,and a pixel-level distribution consistency loss is used to optimize the density map matching process.In addition,there are many background noises that are easily confused with crowd feature in complex scenes.In order to avoid the interference of false positive predictions on subsequent counting and density map estimation,a foreground segmentation module and feature enhancement loss are proposed to adaptively focus the foreground region and increase the contribution of human head features to the counts,so as to suppress background misjudgments.In addition,in order to make the network adapt to the multi-scale pattern of the human head better,up and down sampling operations are performed on each image to be trained to obtain the multi-scale pattern with the same object.Experiments on several datasets show that the proposed method achieves better or competitive results compared with state-of-the-art methods.
Artificial Intelligence
Survey on Evolutionary Recurrent Neural Networks
HU Zhongyuan, XUE Yu, ZHA Jiajie
Computer Science. 2023, 50 (3): 254-265.  doi:10.11896/jsjkx.220600007
Abstract PDF(3002KB) ( 158 )   
References | Related Articles | Metrics
Evolutionary computation utilizes natural selection mechanisms and genetic laws in the process of biological evolution to solve optimization problems.The accuracy and efficiency of the evolutionary recurrent neural network model depends on the optimization effect of parameters and the structures.The utilization of evolutionary computation to solve the problem of adaptive optimization of parameters and structures in recurrent neural networks is a hot spot of automated deep learning.This paper summarizes the algorithms that combine evolutionary algorithms and recurrent neural networks.Firstly,it briefly reviews the traditional categories,common algorithms,and advantages of evolutionary computation.Next,it briefly introduces the structures and characteristics of the recurrent neural network models and analyzes the influencing factors of recurrent neural network perfor-mance.Then,it analyzes the algorithmic framework of evolutionary recurrent neural networks,and the current research development of evolutionary recurrent neural networks from weight optimization,hyperparameter optimization and structure optimization.Besides,other work on evolutionary recurrent neural networks is analyzed.Finally,it points out the challenges and the deve-lopment trend of evolutionary recurrent neural networks.
Document-enhanced Question Answering over Knowledge-Bases
FENG Chengcheng, LIU Pai, JIANG Linying, MEI Xiaohan, GUO Guibing
Computer Science. 2023, 50 (3): 266-275.  doi:10.11896/jsjkx.220300022
Abstract PDF(3488KB) ( 139 )   
References | Related Articles | Metrics
Recently,knowledge base(KB) has been widely adopted to the task of question answering(QA) to provide a proper answer for a given question,known as the KBQA problem.However,knowledge base itself may be incomplete(e.g.KB does not contain the answer to the question,or some of the entities and relationships in the question),limiting the overall performance of existing KBQA models.To resolve this issue,this paper proposes a new model to leverage textual documents for KBQA task by providing additional answers to enhance knowledge base coverage and background information to enhance the representation of questions.Specifically,the proposed model consists of three modules,namely entity and question representation module,document and enhanced-question representation module and answer prediction module.The first module aims to learn the representations of entities from the retrieved subgraph of knowledge base.Then,the question representation can be updated with the fusion of seed entities.The second module attempts to learn a proper representation of the document that is relevant to the given question.Then,the question representation can be further improved by fusing the document information.Finally,the last module makes an answer prediction based on the information of knowledge base,updated question and documents.Extensive experiments are conducted on the WebQuestionsSP dataset,and the results show that better accuracy can be obtained in comparison with other counterparts.
Study on Chinese Named Entity Extraction Rules Based on Boundary Location and Correction
LIU Pan, GUO Yanming, LEI Jun, LAO Mingrui, LI Guohui
Computer Science. 2023, 50 (3): 276-281.  doi:10.11896/jsjkx.220200020
Abstract PDF(1515KB) ( 149 )   
References | Related Articles | Metrics
Compared with English text which is naturally composed of words,Chinese text has no word delimiters,so the combination of Chinese characters is more flexible,and it's more difficult to determine the entity boundaries in Chinese named entity recognition(NER).Current mainstream methods transform the NER task into a sequence labeling task.This paper studies the predicted label sequence under the BIOES tag scheme and calculates the entity boundary accuracy by separately considering the entity head label B or tail label E,which shows that increasing the boundary accuracy can further improve the accuracy of entity recognition.We expand the boundaries of entities with continuous labels,use the label type of the last character of the entity to correct the entity type,and use the word segmentation information to fill in the entity with incomplete labels.Finally,this paper proposes a BIO+ES labeling scheme that adds boundary labels to distinguish non-entity characters at entity boundaries and further improves the performance of Chinese NER.
Chinese Spelling Check Based on BERT and Multi-feature Fusion Embedding
LIU Zhe, YIN Chengfeng, LI Tianrui
Computer Science. 2023, 50 (3): 282-290.  doi:10.11896/jsjkx.220100104
Abstract PDF(2288KB) ( 196 )   
References | Related Articles | Metrics
Due to the diversity of Chinese characters and the complexity of Chinese semantic expressions,Chinese spelling che-cking is still an important and challenging task.Existing solutions usually suffer from the inability to dig deeper into the text semantics and often learn the mapping relationship between incorrect and correct characters through pre-established external resources or heuristic rules when exploiting the unique similarity features of Chinese characters.This paper proposes an end-to-end Chinese spelling checking algorithm model BFMBERT(BiGRU-Fusion Mask BERT) that incorporates multi-feature embedding of Chinese characters.The model first uses a pre-training task combining confusion sets to make BERT learn Chinese spelling error knowledge.It then employs a bi-directional GRU network to capture the probability of error for each character in the text.Furthermore,it applies this probability to compute a fusion embedding incorporating semantic,pinyin,and glyph features of Chinese characters.Finally,it feeds this fusion embedding into a mask language model in BERT to predict correct characters.BFMBERT is evaluated on the SIGHAN 2015 benchmark dataset and achieves an F1 value of 82.2,outperforming other baseline models.
Employing Gated Mechanism to Incorporate Multi-features into Chinese Event Coreference Resolution
HUAN Zhigang, JIANG Guoquan, ZHANG Yujian, LIU Liu, LIU Shanshan
Computer Science. 2023, 50 (3): 291-297.  doi:10.11896/jsjkx.220700146
Abstract PDF(2960KB) ( 124 )   
References | Related Articles | Metrics
Event coreference resolution is the basis of many natural language processing tasks,aiming to identify event mentions in text that refer to the same real event.Since Chinese grammar is much more complex than English,the method of capturing English text features is not effective in Chinese event corefe-rence resolution.To solve the within-document Chinese event corefe-rence,a gated mechanism neural network(GMNN) is proposed.In view of Chinese characteristics with subject omission and loose structure,event attributes are introduced as symbolic features.On this basis,a novel gated mechanism is proposed,which fine-tunes the symbolic feature vector,filters the noise in the symbolic features,extracts useful information in a specific context,and improves the coreference events recognition rate.Experimental results on the ACE2005 Chinese dataset show that the perfor-mance of GMNN improves by 2.66,which effectively improves the effect of Chinese event coreference resolution.
Multimodal Sentiment Analysis Based on Adaptive Gated Information Fusion
CHEN Zhen, PU Yuanyuan, ZHAO Zhengpeng, XU Dan, QIAN Wenhua
Computer Science. 2023, 50 (3): 298-306.  doi:10.11896/jsjkx.220100156
Abstract PDF(3144KB) ( 185 )   
References | Related Articles | Metrics
The goal of multimodal sentiment analysis is to achieve reliable and robust sentiment analysis by utilizing complementary information provided by multiple modalities.Recently,extracting deep semantic features by neural networks has achieved remarkable results in multimodal sentiment analysis.But the fusion of features at different levels of multimodal information is also an important part in determining the effectiveness of sentiment analysis.Thus,a multimodal sentiment analysis model based on adaptive gating information fusion(AGIF) is proposed.Firstly,the different levels of visual and color features extracted by swin transformer and ResNet are organically fused through a gated information fusion network based on their contribution to sentiment analysis.Secondly,the sentiment of an image is often expressed by multiple subtle local regions due to the abstraction and complexity of sentiment,and these sentiment discriminating regions can be located accurately by iterative attention based on past information.The latest ERNIE pre-training model is utilized to solve the problem of Word2Vec and GloVe's inability to handle the word polysemy.Finally,the auto-fusion network is utilized to “dynamically” fuse the features of each modality,solving the pro-blem of information redundancy caused by the deterministic operation(concatenation or TFN) to construct multimodal joint representation.Extensive experiments on three publicly available real datasets demonstrate the effectiveness of the proposed model.
Sentiment Analysis of Chinese Short Text Combining Context and Dependent Syntactic Information
DU Qiming, LI Nan, LIU Wenfu, YANG Shudan, YUE Feng
Computer Science. 2023, 50 (3): 307-314.  doi:10.11896/jsjkx.211200189
Abstract PDF(2301KB) ( 196 )   
References | Related Articles | Metrics
Dependency parsing aims to analyze the syntactic structure of sentences from the perspective of linguistics.Existing studies suggest that combining such graph-like data with graph convolutional network(GCN) can help model better understand the text semantics.However,when dealing with dependency syntactic information as adjacency matrix,these methods ignore the types of syntactic dependency tags and the word semantics related to the tags,which makes the model unable to capture the deep emotional features.To solve the preceding problem,this paper proposes a Chinese short text sentiment analysis model CDSI(context and dependency syntactic information).This model can use BiLSTM(bidirectional long short-term memory) network to extract the context semantics of the text.Moreover,a dependency-aware embedding representation method is introduced to mine the contribution weights of different dependent paths to the sentiment classification task based on the syntactic structure.Then the GCN is used to model the context and dependent syntactic information at the same time,so as to strengthen the emotional features in the text representation.Based on SWB,NLPCC2014 and SMP2020-EWEC datasets,experimental results show that CDSI can effectively integrate the semantic and structural information in sentences,which achieves good results in both the Chinese short text sentiment binary classification and multi-classification tasks.
Chinese Argumentative Writing Quality Evaluation Based on Multi-perspective Modeling
HE Yaqiong, JIANG Feng, CHU Xiaomin, LI Peifeng
Computer Science. 2023, 50 (3): 315-322.  doi:10.11896/jsjkx.220100137
Abstract PDF(1859KB) ( 155 )   
References | Related Articles | Metrics
Automated essay scoring is a task that replaces manual grading for students’ essays,where rich semantics,rigorous organization,and reasonable logic are important considering factors.Most previous studies only consider the semantics or organization of the essay from a single perspective,lacking considering higher-level factors such as logic.Therefore,this paper proposes a multi-perspective evaluation framework(MPE) to more objective and reliable evaluate the essay from semantics,organization,and logic.MPE first utilizes the pre-trained model to encode sentence and obtain three levels semantic information to evaluate the essay's semantic expression.Then,it combines sentence function identification and paragraph function identification to evaluate the essay′s organization.Moreover,MPE evaluates the essay's logic by calculating the coherence between paragraphs.Finally,the framework scores the essay by integrating these three evaluation perspectives.Experimental results show that the proposed multi-perspective evaluation framework can effectively score the essays at various qualities,outperforming all the baselines.
Real-time Trajectory Planning Algorithm Based on Collision Criticality and Deep Reinforcement Learning
XU Linling, ZHOU Yuan, HUANG Hongyun, LIU Yang
Computer Science. 2023, 50 (3): 323-332.  doi:10.11896/jsjkx.220100007
Abstract PDF(4622KB) ( 197 )   
References | Related Articles | Metrics
Real-time collision avoidance in dynamic environments is a challenge in trajectory planning of mobile robots. Focusing on environments with variable number of obstacles,this paper proposes a real-time trajectory planning algorithm,Crit-LSTM-DRL,based on long short-term memory(LSTM) and deep reinforcement learning(DRL). First,it predicts the time to the occurrence of a collision between an obstacle and the robot based on their states,and then computes the collision criticality of each obstacle with respect to the robot. Second,it generates the obstacle sequence based on the collision criticality and abstracts a fixed-dimension vector by LSTM to represent the environment. Finally,the robot state and the extracted vector are concatenated as the input of the DRL's value network to compute the value with respect to the system state. At any instant,for each action,it predicts the value of the next state based on the LSTM and DRL models and then the value of the current state; hence,the action generating the maximal value of the current state is selected to control the robot. To evaluate the performance of Crit-LSTM-DRL,it is first trained in three different environments and obtain three models: the model trained in the environment with 5 obstacles,the model trained in the environment with 10 obstacles,and the model trained in the environment with variable number of obstacles(1~10). The models then are tested in various environments containing different number of obstacles. To further investigate the effects of the interaction between an obstacle and the robot,this paper also takes the joint state of an obstacle and the robot as the state of the obstacle and trains another three models in the above training environments. Experimental results show the effectiveness and efficiency of Crit-LSTM-DRL.
Information Security
Backdoor Attack on Deep Learning Models:A Survey
YING Zonghao, WU Bin
Computer Science. 2023, 50 (3): 333-350.  doi:10.11896/jsjkx.220600031
Abstract PDF(2563KB) ( 281 )   
References | Related Articles | Metrics
In recent years,artificial intelligence represented by deep learning has made breakthroughs in theories and technologies.With the strong support of data,algorithms and computing power,deep learning has received unprecedented attention and has been widely used in various fields,bringing great improvements to the corresponding fields.With the wide application of deep learning technology in various fields including security critical ones,the security issue of deep learning has attracted more and more attention.Researchers have found many security risks in deep learning systems.In terms of the security of deep learning models,researchers have extensively explored the new attack paradigm of backdoor attack.Backdoor attack can threaten deep learning models throughout their whole life cycle.A large number of researchers have proposed series of attack scheme from different angles.This paper takes the security threats of deep learning system as a starting point,introduces the current attack paradigms.On this basis,it gives the back-ground and principle of backdoor attack,distinguishes the similar attack paradigms such as adversarial attack and data poisoning attack,then continues to elaborate on the attack principle and outstanding features of the classic methods of backdoor attack to date.According to the working principle,the attack schemes are divided into data poisoning based attack and model poisoning based attack and others,the paper systematically summarizes them and clarify the advantages and disadvantages of current research.Then,this paper surveys the state-of-the-art works of backdoor attack against various typical applications and popular deep learning paradigms,which further reveal the threat of backdoor attack towards deep learning models.Finally,this paper summarizes the research work on applying backdoor attack characteristics to positive applications and explores the current challenges of backdoor attack,as well as discusses future research directions worthy of in-depth exploration,aiming to provide guidance for the follow-up researchers to further promote the development of backdoor attack and security of deep learning.
Survey on Membership Inference Attacks Against Machine Learning
PENG Yuefeng, ZHAO Bo, LIU Hui, AN Yang
Computer Science. 2023, 50 (3): 351-359.  doi:10.11896/jsjkx.220100016
Abstract PDF(1995KB) ( 179 )   
References | Related Articles | Metrics
In recent years,machine learning has not only achieved remarkable results in conventional fields such as computer vision and natural language processing,but also been widely applied to process sensitive data such as face images,financial data and medical information.Recently,researchers find that machine learning models will remember the data in their training sets,making them vulnerable to membership inference attacks,that is,the attacker can infer whether the given data exists in the training set of a specific machine learning model.The success of membership inference attacks may lead to serious individual privacy leakage.For example,the existence of a patient's medical record in a hospital's analytical training set reveals that the patient was once a patient there.The paper first introduces the basic principle of membership inference attacks,and then systematically summarizes and classifies the representative research achievements on membership inference attacks and defenses in recent years.In particular,how to attack and defend under different conditions is described in detail.Finally,by reviewing the development of membership inference attacks,this paper explores the main challenges and potential development directions of machine learning privacy protection in the future.
Efficiently Secure Architecture for Future Network
YANG Xin, LI Hui, QUE Jianming, MA Zhentai, LI Gengxin, YAO Yao, WANG Bin, JIANG Fuli
Computer Science. 2023, 50 (3): 360-370.  doi:10.11896/jsjkx.220600265
Abstract PDF(5496KB) ( 160 )   
References | Related Articles | Metrics
Traditional IP-based Internet offers an end-to-end data transport service and has developed rapidly in the past half-century.However,serious security incidents emerged from attacks based on traditional networks.Traditional security mechanisms(e.g.,firewalls,intrusion detection systems) enhance security.However,most of them only provide some remedial strategies rather than solve the address-security problem radically due to the lack of change in network design.The overall in-depth security of the networked system cannot be guaranteed without a fundamental change.In order to meet the development requirements of the next generation of an endogenous security network,one of the future networks,the multi-identifier network(MIN),is introduced as our research background.This paper proposes an efficient scheme in hieratical architecture that provides comprehensive protection by addressing the security aspects pertaining to the network and application layers.At the network layer,the proposed architecture develops a multi-identifier routing scheme with embedded identity-based authentication and packet signature mechanisms to provide data tamper-resistance and traceability.At the application layer,the proposed architecture designs a mimic defensive scheme combined with weighted network centrality measures.This scheme focuses on protecting the core components of the whole network to improve the service's robustness and efficiently resist potential attacks.This paper tests and evaluates the proposed scheme from a theoretical and practical perspective.An analytical model is built based on the random walk for theoretical evaluation.In experiments,the proposed scheme is developed in MIN as MIN-VPN.Then considering IP-VPN as a baseline,anti-attack tests are conducted on IP-VPN and MIN-VPN.The results of theoretical evaluations and experiments show that the proposed scheme provides excellent transmission performance and successful defense against various TCP/IP-based attacks with acceptable defensive cost,demonstrating this security mechanism's effectiveness.In addition,after long-period penetration testing in three international elite security contests,the proposed method is effectively immune to all TCP/IP-based attacks from thousands of professional teams,thus verifying its strong security.
Network Equipment Anomaly Detection Based on Time Delay Feature
CUI Jingsong, ZHANG Tongtong, GUO Chi, GUO Wenfei
Computer Science. 2023, 50 (3): 371-379.  doi:10.11896/jsjkx.211200280
Abstract PDF(3202KB) ( 168 )   
References | Related Articles | Metrics
With the rapid development of the Internet,the security of network equipment has received extensive attention.Aiming at the problems of that the existing network equipment anomaly detection technology is destructive and difficult to detect,the paper uses the packets delay spent by the network equipment to transmit and process data packets as the detection basis,and proposes an anomaly detection scheme based on delay characteristics.The proposed scheme adopts side channel analysis,and it does not need to upgrade the equipment's software or hardware.It has the characteristics of non-intrusive and easy to implement.Firstly,the method uses the high-precision timing technology time stamp machine to collect the time delay information,and uses the genetic algorithm to extract the peak position feature of the delay distribution.Secondly,to solve the problem of the imbalance of data set,the method uses one-class support vector machine algorithm to construct anomaly detection algorithm.Finally,the validity of the method is verified by building an experimental platform,and the experimental results are evaluated.Experimental results show that the proposed method is feasible and effective.
Semi-supervised Network Traffic Anomaly Detection Method Based on GRU
LI Haitao, WANG Ruimin, DONG Weiyu, JIANG Liehui
Computer Science. 2023, 50 (3): 380-390.  doi:10.11896/jsjkx.220100032
Abstract PDF(2137KB) ( 181 )   
References | Related Articles | Metrics
Intrusion detection system(IDS) is a detection system that can issue an alarm when a network attack occurs.Detecting unknown attacks in the network is a challenge that IDS faces.Deep learning technology plays an important role in network traffic anomaly detection,but most of the existing methods have a high false positive rate and most of the models are trained using supervised learning methods.A gated recurrent unit network(GRU)-based semi-supervised network traffic anomaly detection me-thod(SEMI-GRU) is proposed,which combines a multi-layer bidirectional gated recurrent unit neural network(MLB-GRU) and an improved feedforward neural network(FNN).Data oversampling technology and semi-supervised learning training method are used to test the effect of network traffic anomaly detection using binary classification and multi-classification methods,and NSL-KDD,UNSW-NB15 and CIC-Bell-DNS-EXF-2021 datasets are used for verification.Compared with classic machine learning mo-dels and deep learning models such as DNN and ANN,the SEMI-GRU method outperforms the machines lear-ning and deep learning methods listed in this paper in terms of accuracy,precision,recall,false positives,and F1 scores.In the NSL-KDD binary and multi-class tasks,SEMI-GRU outperforms other methods on the F1 score metric,which is 93.08% and 82.15%,respectively.In the UNSW-NB15 binary and multi-class tasks,SEMI-GRU outperforms the other methods on the F1 score,which is 88.13% and 75.24%,respectively.In the CIC-Bell-DNS-EXF-2021 light file attack dataset binary classification task,all test data are classified correctly.
Ransomware Early Detection Method Based on Deep Learning
LIU Wenjing, GUO Chun, SHEN Guowei, XIE Bo, LYU Xiaodan
Computer Science. 2023, 50 (3): 391-398.  doi:10.11896/jsjkx.220200182
Abstract PDF(2288KB) ( 166 )   
References | Related Articles | Metrics
In recent years,ransomware is becoming increasingly prevalent,causing serious economic losses.Since files encrypted by ransomware are difficult to recover,how to timely and accurately detect ransomware is a hot point nowadays.To improve the timeliness and accuracy of ransomware detection,this paper analyzes the behavior of ransomware family and benign software in the early stage of operation and proposes a ransomware early detection method based on deep learning(REDMDL).REDMDL takes a certain length of application programming interface(API) sequence that is obtained by software running at the initial stage as input,combines word vector and position vector to vectorize the collected API sequence,and then constructs a convolutional neural network-long short term memory(CNN-LSTM) neural network model for early detection of ransomware.Experimental results show that REDMDL can accurately determine whether the software is ransomware or benign within seconds after it star-ting to run.