مجله رایانش نرم و فناوری اطلاعات
سال یازدهم شماره 1 (بهار 1401)

The main purpose of this study is to represent the semantic word vectors with low dimensions, explicitly. The problem of finding a limited number of meaningful basis vectors for producing explicit semantic word vectors must be solved in such a way that a large accuracy drop is not caused by reducing the dimensions. In this study, we represent a hybrid approach to finding meaningful basis vectors. First, we obtain N basis vectors using the proposed methods 1- The criterion of word similarity-to-word frequency ratio, 2- Feature selection method based on comparison of distance matrices, 3- Binary weighting method based on PSO algorithm. Then, to take advantage of the expertise of methods 1 and 2 to the same extent, we obtain the first combined basis vectors by combining half of the basis vectors obtained by the criterion of word similarity-to-word frequency ratio with half of the basis vectors selected by the feature selection method. In the next step, we obtain the common context words that have a weight "1" as the common basis vectors produced by the binary weighting method. In the next step, we add the common context words with a weight "1" obtained using the BPSO method to the first combined basis vectors obtained from word similarity-to-word frequency ratio and the feature selection methods. Thus, the second combined basis vectors are obtained, which are meaningful, and each basis vector is equivalent to an informative context word. Therefore, the explicit word vectors produced by meaningful basis vectors can be interpreted. We train the proposed approach using the UkWaC corpus and evaluate it using the word similarity task. Both first and second combined basis vectors improve accuracy. The increase in accuracy is greater in the first combined basis vectors. The evaluation results of explicit word vectors obtained with the first basis vectors show that despite the reduction of word vector dimensions from 5000 to 1511, the Spearman correlation coefficient on MEN, RG-65, and SimLex-999 test sets is increased by 2.47%, 7.39%, and 0.52%, respectively.

Semantic Textual similarity is one of the subtasks of natural language processing that has attracted extensive rese arch in recent years. Measuring semantic similarity between words, sentences, paragraphs, and documents plays an important role in natural language processing and computational linguistics. Semantic similarity of texts is used in question-answering systems, fraud detection, machine translation, information retrieval and etc. Semantic similarity means calculating the degree of similarity between two textual documents, paragraphs or sentences, which are presented in both monolingual and cross lingual forms. In this article, by using the parallel corpus, for the first time, the cross lingual model of semantic similarity for Persian-English sentences is presented, and then we test and compare our model with the Multilingual BERT model. The results show that by using parallel corpuses, the quality of sentence embedding in two different languages can be improved. Pearson correlation criterion based on cosine similarity between sentence's vector of multilingual Bert has increased from 65% to 73.77% by the proposed method. The proposed method was also tested on the Arabic-English language pair, and the results show that the proposed method is superior to the multilingual Bert.

Software Defined Network (SDN) has become a popular model for centralized control and management in many modern network scenarios. However, for large data centers with hundreds of thousands of servers and several thousand switches, a single controller mode causes the system to suffer from lack of scalability and reliability. The use of distributed architecture can improve system performance, but the main limitation of this work is the static mapping between the switch and the controller, which may lead to load imbalances in the controllers. The use of multiple controllers distributed in SDN has been used to improve scalability and reliability, where each controller manages a fixed partition of the network. In the proposed method, a balanced controller (BalCon) is used as a scheme to dynamic migrate the SDN switch to achieve load balance between SDN controllers with low migration cost. The decision to migrate is made by the monitor based on network traffic. To balance load in the software-based network, a staged switch migration strategy is used and the target controller is selected by the gray wolf algorithm. The results show that the response time of the proposed method is 13% better than the BalCon method. Also, by increasing the sending rate to 8.33%, the controller's throughput improves compared to the BalCon method. Therefore, the load balance of the proposed method has a better performance than previous works.

In recent decades, with the popularity and application of cloud computing, significant changes have taken place in communications and technologies. Most service-level service providers are offering applications and developing their own hardware infrastructure as service improvements. Since cloud computing provides different resources to these providers, of course, the cost of access, speed, and other important parameters have led to a significant response, but an important point has given the increase and volume of requests from stakeholders. It has led to challenges in providing services at the service level. Therefore, scheduling and allocating resources to requests made with low-cost horizons and completion time has become a necessity, and service providers and stakeholders seek to receive the best possible service with high efficiency, and this has led to extensive research in this area. In this research, a model for scheduling and resource allocation considering priority and SLA in cloud computing is presented. In fact, the proposed model of several different levels of access has been developed to achieve the main goal of the research, which is the optimal scheduling and allocation of resources to the requests made. In the proposed method, using the RR algorithm and the technique of weighting requests and online review of virtual machines, the best possible source for proposal allocation and scheduling has been identified. The method presented in the form of a dynamic and executable model with the help of the Cloudsim simulation tool and in terms of Makespan, cost, and speedprocess have been compared and analyzed with several similar methods. The results obtained from the simulation performed by applying different scenarios indicate average processing speed around 2.15, and average Makespan is reduced at 8.68s by new method than similarity methods. Also, the rang of cost has not big change.

In the last decade, several convolutional networks have been developed for the semantic segmentation, which have shown excellent performance in recognizing and labeling objects in images. Most of these networks involve large-scale architectures that can detect tens or hundreds of predefined classes. With the exception of fully convolutional networks, most applications use architectures that, after convolutional layers, use a common classifier to classify the extracted features. In this paper, the method of converting a network, which as classifier, has two flatten and dense layers (fully connected), to a fully convolutional network is described. The main advantage of this method is the ability to work on inputs of variable size and produce an output map instead of a number, which is the advantage of fully convolutional networks. Newer models of the Deep Learning area generally use training images in which areas of interest are determined by masks; but in the proposed method only labeled images are given to the network. The details of the proposed method are expressed in the form of a new problem of classification of boards with calligraphy of Shekasteh-Nastaliq and Suls, and classification of apple leaf diseases (as two-class problems) and the problem of identifying hand written Persian digits. For this purpose, first a convolutional network with the last fully connected layer is designed and trained for square images. Then a new fully convolutional model is defined based on the previous model and the weights of the previous model are fed to the new model. The only difference between the two models is in the last layer, but the new model will be able to work on input images of any size. Experimental results show the efficiency of the proposed approach.

The construction of classification models is widely used in data mining. There are concerns about the privacy of data owners because of the need to collect data to build models. In this paper, a Naïve Bayes classification model construction plan is presented, which performs the model construction operation with the participation of data owners and without the need to collect the original data. Instead of collecting data, the scheme uses encryption of bit strings from counting without disclosing the original data to perform the process of creating the Naïve Bayes model. This design allows the model to be built with appropriate performance without the need for trust in a third party with a minimum number of encryptions, so that in terms of time complexity, up to 87% improvement in time cost can be observed. In addition, memory consumption has not increased significantly when compared to designs that use encryption operations.

In wireless communications which are widely applicable nowadays, the channel variation over time and frequency is a major problem. Therefore, the channel estimation and compensation are critical challenges in these systems. The OFDM technique is noteworthy for wireless channels due to robustness, high efficiency and consequently high data rate. Pilot-aided channel estimation is a promising approach in many OFDM systems, including mobile networks, digital audio and video broadcasting standards. The number and the arrangement of pilots in the time-frequency grid affect the accuracy and the period of the channel estimation procedure. Practical OFDM systems utilize fixed pilot patterns that are designed based on typical assumptions of the fading channel. Thus, unnecessary pilot overhead is imposed when the channel is close to a flat slow fading channel, and the performance degrades in the worst condition of the channel. To remedy this issue, adaptive pilot arrangement based on the channel response is introduced in the previous researches. Indeed, finding the best pilot arrangement according to the channel response is an optimization problem and the adaptive approach is exploited to find its answer. In this paper, we develop a novel method for adaptive pilot arrangement based on Kalman filter which is combined with the channel estimation algorithm. Therefore, the position of the pilots and the channel response are jointly estimated in the proposed method. This method is simulated for various channel types and its performance is evaluated by bit error rate (BER) and pilot overhead criteria. Simulation results illustrate that the proposed method obtains the same performance as the previous methods by fewer pilots. This achievement leads to significant increase in the capacity of data transfer.