While a person designs a choice question with multiple options, he or she requires some incorrect options as well as correct option. And incorrect options need to be looked like correct so that the correct option cannot be easily selected out. Question answering(QA) can hardly provide incorrect options because it only focuses on searching correct answer. This paper gives a method to generate incorrect options for a certain question and a given correct option. Confusion degree is proposed to describe the degree of an incorrect option being mistakenly considered as correct. Experiment shows that the method can effectively generate incorrect and confusing options for a question. The method can help design choice questions in an examination.
This paper verifies a probable link between collaboration link and citation link through experiments on scientific papers. The result shows that scholars tend to cooperate with scholars with lower citation numbers. More than 70% of the cooperation between high-citation scholars and low-citation scholars comes from the same institution.
This paper surveys research on the Probabilistic Resource Space Model (P-RSM), which is a probabilistic, multidimensional classification space model. It can effectively manage various uncertain resources with a complete theoretical basis. This paper introduces the basic definition, paradigms, operations, constraints of this model, and discusses its usefulness through application instances.
Recommending appropriate collaborators to researchers can promote their research. In many cases, however, it is difficult for researchers to find proper collaborators from large number candidates. This paper proposes a scientific collaborator recommendation approach based on the semantic link networks, where nodes are authors, papers and interests indicated by keywords, and semantic links are write links, cite links, and contain links between these semantic nodes. Five semantic paths on the semantic link networks are proposed for deriving future collaboration between authors. Experiments on three datasets of scientific journal papers show that our method achieves good performance in predicting future collaborators. Comparing the combinations of five semantic paths reaches the following results: (1) co-author relationship, keyword information, and citation relationship play an important role in finding appropriate collaborators; and, (2) combining all the five semantic paths can get the best results on collaborator recommendation task.
With the rapid development of natural language and the implementation of the Wisdom Court, intelligent judicial assistants has become a new application of natural language processing in the judicial field. The text classification method based on word vector and deep neural network implements statistical classification of judicial documents, but it can not achieve the inherent logical interpretation of judicial cases. A method of extracting semantic logic tree from judicial case texts is proposed, and event tree can be interpreted by deep forest. Judicial documents are divided into several sub tree fragments by sentence segmentation, and each sub-tree fragment is analyzed by dependency syntax to obtain core subject-predicate-object triples. TF-IDF algorithm is used to calculate the weights of triples, and get the core sub-event sequence, and use pruning algorithm to construct the max heap. The designed triple encoding algorithm realizes max heap vectorization of event tree, and embedded deep forest algorithm to realize classification discrimination of judicial text event tree. The experimental results show that the proposed event tree construction method combined with the deep forest algorithm can greatly improve the logical interpretation and accuracy of judicial text.
Image retrieval technology has made great breakthroughs in the development of computer vision. Feature extraction is crucial to image retrieval as a good method not only brings convenience to image recognition, but also improves the performance of image retrieval system. Many traditional methods only extract shallow features like the statistics of brightness, color components and texture measures, which miss the contextual semantic information and cannot behave well in image retrieval. However, feature extraction based on deep learning is able to obtain better features with semantic information. Therefore, much work has shown that integrating neural network with hash codes performs better in retrieval tasks. Encouraged by this, we propose an image hash retrieval algorithm by optimizing structures in deep layer aggregation network (DlaNet). This model is mainly composed of an improved net called DlaNet-V and a four-valued hash code scheme. The DlaNet-V is optimized on the basis of DlaNet, improving the efficiency. Meanwhile, the binary hash codes are switched into the quaternary hash codes to make the model more robust and efficient. Experiments are conducted on the CIFAR-10 data set and a medical device image data set collected by authors. Results show that the image retrieval based on DlaNet-V and quaternary hash code is more accurate and stable.
A reader sometimes requires a text containing more specific or general content based on current understanding on a certain field or topic. Traditional text recommendation systems barely meet the requirement because they recommend texts according to the relevance of semantics or topics. This paper proposes a text recommendation method that can provide texts with more specific or general content based on a user's current understanding. Experiments show that the method is effective and useful in the kind of text recommendation.
Unlike the instrumental need of language, linguistic humanity of language means that languae, like Chinese verses, always show certain emotion, reflect attitude toward thing and convey creator's characters.To model and visualize such complicated and cross Humanism, Knowledge Graph(KG) is modified to adapt the full-relation Chinese verses and leading-in friendly visualization.Different from conventional Ml or rule-based construction,we adopt the group-intelligence construction with entity adding task, linking adding task and knowledge triple validation task to constructe the KG by alocating task destributed model.While using group-intelligence construction may lead to multi-answer and irrelevant-answer case, socring mechanism with user-consideration and group consideration is definded.This research shows construction based on the group intelligence can optimize the KG and represent different congnize of Chinese Verses.
Sentiment lexicons play a vital role in the field of sentiment classification. The existing sentiment lexicon has problems such as limited coverage and poor adaptability in the field. Building an sentiment lexicon with large coverage and strong adaptability in the field has become a challenge in this field. This paper proposes a new method for selecting seed words. Firstly, the seed words are manually selected based on the general lexicon. Then, the word vectors are trained on the corpus by the seed words selected artificially. Finally, the expanded seed words are obtained. Obtaining sentiment polarity by calculating the similarity between the seed words and the candidate sentiment words, constructing propagation map and propagation matrix to construct sentiment lexicon. The experimental results show that the method can obtain higher accuracy and better robustness compared with the baseline method.
In view of most of sentiment analysis texts are too short to get enough textual features, a method of bidirectional Long Short-Term Memory using tone (Word, Character and Tone model based on Bidirectional Long Short-Term Memory, WCT-Bi-LSTM) was proposed. Distinguished from the general method of sentiment analysis only taking word as the feature, the model also used character and tone features as input to enrich the characteristics of the text. After that, the model integrated the deep semantic meaning of word, character and tone. It could better grasp the emotion of the text and improve the accuracy of sentiment classification. The experimental results show that, compared with the model which does not integrate tone, the accuracy of the proposed model is increased by 1.2% and 0.9%on two experimental datasets respectively, which proves that the proposed method can effectively improve the accuracy of sentiment classification.
With rapid expansion of scientific papers, making a survey from a large collection of papers on a given research issue or domain becomes more and more important for researchers. This paper proposes a template-based framework for automatically generating survey paper. It allows users to compose a template tree as a syllabus for the required survey. Each tree node corresponds to a section to be composed in the survey therefore the whole tree defines the section structure of the survey. The template consists of two types of nodes, dimension node and topic node, which filter contents of papers. A recursive procedure along the survey generation template tree paths is conducted to process documents, rank sentences, and compose sections. We apply the approach to generating the survey of the reference papers of a survey paper and compare the result with the survey paper. Experiments show improvement over several baseline methods.
There has been much interest in link prediction research with significant studies on how to predict missing links or future links in a network based on observed information. However, the key solution to tackle the link prediction problem is how to measure the similarity between the nodes in a network with higher accuracy. Several methods have been proposed that applies the similarity between nodes to estimate their proximity in the network. In this paper, an efficient link prediction algorithm that predicts relationships between links using the network structure is proposed, which uses common neighbors in addition to the degree distribution of the nodes to estimate the possibility of the presence of a link between two nodes in a network based on local information. Extensive experiments are carried out and compared with 10 standard similarity-based methods using 7 real-world datasets. The experimental results show that our proposed method has higher prediction accuracy compared with most of the local information based methods like the Common Neighbor and Preferential Attachment. It is also competitive with the quasi-local indicators such as LP and global indicators like Katz, with a lower computational complexity.
The idea of Word Embedding is based on the semantic distribution hypothesis of the linguist Harris (1954), who believes that words of the same semantics are distributed in similar contexts. Learning of vector-space word embeddings is a technique of central importance in natural language processing. In recent years, cross-lingual word vectors have received more and more attention. Cross-lingual word vectors enable knowledge transfer between different languages, the most important It is this transfer that can take place between resource-rich and low resource languages. This paper uses Tibetan and Chinese Wikipedia corpus to train monolingual word vectors, mainly using the fastText word vector training method, and the two monolingual word vectors are analyzed by CCA correlation, thus obtaining Tibetan-Chinese cross-lingual word vectors. In the experiment, we evaluated the resulting word representations on standard lexical semantic evaluation tasks and the results show that this method has a certain improvement on the semantic representation of the word vector.
Today, cloud computing has become an advanced form of distributed computing, grid computing, utility computing, and virtualization. Efficient task scheduling algorithms help to reduce the number of virtual machines used, thus reducing costs and improving stability. To solve the problem of cloud computing task scheduling, an improved particle swarm optimization (IPSO) task scheduling method is proposed based on the traditional PSO algorithm. Firstly, this paper describes the mathematical model of cloud computing task scheduling and the basic principle of particle swarm optimization. On this basis, the random method is used to generate the initial population definition appropriateness function, the indirect coding method is used to encode the resources, and the time-varying method is used to adjust the inertia weight. In the position update, according to the inertia weight w, the individual optimal value Pbest or the group optimal value Gbest is legalized to determine the update method of the particle velocity and position, thereby increasing the degree of discretization of the PSO algorithm. The simulation test on the CloudSim platform shows that the scheduling strategy is effective and efficient. Experimental results demonstrate that the proposed method obtains better scheduling results. Thereby controlling global search and local search, try to avoid falling into local optimum.
Currently, information distribution based on publish/subscribe is widely used in battlefield situation information sharing, where the theme-oriented or content-oriented ideas are mainly adopted. However, information requirements of the same combat unit are different when executing different missions. And traditional publish/subscribe based information distribution is not competent enough to meet the dynamic information requirements of combat units that are closely related to missions. We propose a mission-oriented dynamic distribution method of battlefield situation information with a feedback mechanism. Where a structured general concept model is designed to realize a consistent representation of heterogeneous battlefield information from different information systems, based on which, a unified information space is built. A matching strategy between missions and classified battlefield situation information based on rules is designed to realize mission-oriented customization of information. An information feedback mechanism based on content and semantics is proposed to realize information correctness and integrity verification based on the applied-logic of missions, as well as feedback information generating. Theoretical analysis shows that, feedback-based information distribution can provide required information for different combat units dynamicly, at the same time, it is conductive to achieve mission-oriented targeted acquisition of situation information and to realize the complement of battlefield situation information space.
This paper proposes a clause-based extractive summarization algorithm by ranking and extracting semantic clauses from the original document. Discourse structure relation is useful for identifying semantically important parts of the source document. We segment the document into clauses and evaluate the importance of clauses based on semantic relations, and then, rank and extract them coarsely, and utilize graph rank to refine the extracted clauses. This way can create a more concise summary with more information and less redundancy. Research reach the following results: 1) compared with the other summarization algorithms on different granularity, the clause-based summarization achieves higher recall score; and, 2) different discourse relations have different importance.
The Resource Space Model is a systematic theory and method for managing the contents of resources by a multi-dimensional classification space. Different dimensions represent different classification methods. Each dimension can be regarded as a classification tree. Design of Resource Space is mainly based on domain knowledge and experience of designer. This paper implements an approach to automatic discovering dimensions from texts (taking scientific papers as experimental data). It consists of four stages: 1) representing each text as a collection of tuples in the form of (word, value); 2) constructing an initial dimension; 3) constructing the hierarchical dimensions based on the initial dimension; and 4) post-processing of outlier texts. The experimental results show that our model achieves a significant improvement of performance on classifying texts.
Over the years, the increasing development of location acquisition devices have generated a significant amount of spatio-temporal data. This data can be further analysed in search for some interesting patterns, new information, or to construct predictive models such as next location prediction. The goal of this paper is to contribute to the future research and development of group pattern discovery algorithms from spatio-temporal data by providing an insight into algorithms design in this research area which is based on a comprehensive classification of state-of-the-art models. This work includes static, big data as well as data stream processing models which to the best of authors' knowledge is the first attempt of presenting them in this context. Furthermore, currently available surveys and taxonomies in this research area do not focus on group pattern mining algorithms nor include the state-of-the-art models. The authors conclude with the proposal of a conceptual model of Universal, Streaming, Distributed and Parameter-light (USDP) algorithm that addresses current challenges in this research area.
Datasets are the important part of vehicle re-identification (re-id) research. The dataset which represents real world environment is crucial to vehicle re-id steps such as learning visual features, vehicle detection, examining performance of vehicle re-id algorithms, and so on. Often vehicle re-id datasets lacks in this context. In this paper, firstly, we investigate the vehicle re-id datasets bias problem using deep CNN model inception-v3 (Dataset classification). Dataset classification results indicates that current available vehicle re-id datasets are highly biased. Secondly, we present novel data augmentation technique to mitigate this issue by inserting additional type of variability in training set. Extensive experimental results shows that our approach can be helpful to minimize training set bias. Consequently, cross dataset vehicle re-id performance improves.
Today, to share business data among multiple target database systems is a challenging problem, in particular, in cases of companies combinations, business integration and commercial cooperation. To solve this problem, this paper proposes a meta-database system approach to business data sharing. Our new methodology is to build a meta-database system to manage multiple target databases such that end-users can share business data by cross-search over the target databases but each target database can be managed in its own traditional way. Moreover, a new target database can be easily added into the system and an old one can also be easily deleted from the system. The major advantage of this approach is that it can provide the fastest means and lowest cost for business data sharing. Therefore, it is very effective for those business cases that require urgent data sharing. By providing a general-purpose package and related tools, this methodology can be easily applied in various areas. This is the first attempt to solve the business data sharing problem by using a meta-database system.
