The number of open source language models that can produce Turkish is increasing day by day, as in other languages. In order to create the basic versions of such models, the training of multilingual models is usually continued with Turkish corpora. The alternative is to train the model with only Turkish corpora. In this study, we first introduce the cosmosGPT models that we created with this alternative method. Then, we introduce new finetune datasets for basic language models to fulfill user requests and new evaluation datasets for measuring the capabilities of Turkish language models. Finally, a comprehensive comparison of the adapted Turkish language models on different capabilities is presented. The results show that the language models we built with the monolingual corpus have promising performance despite being about 10 times smaller than the others.

Abstract Integrating adversarial machine learning with question answering (QA) systems has emerged as a critical area for understanding the vulnerabilities and robustness of these systems. This article aims to review adversarial example-generation techniques in the QA field, including textual and multimodal contexts. We examine the techniques employed through systematic categorization, providing a structured review. Beginning with an overview of traditional QA models, we traverse the adversarial example generation by exploring rule-based perturbations and advanced generative models. We then extend our research to include multimodal QA systems, analyze them across various methods, and examine generative models, seq2seq architectures, and hybrid methodologies. Our research grows to different defense strategies, adversarial datasets, and evaluation metrics and illustrates the literature on adversarial QA. Finally, the paper considers the future landscape of adversarial question generation, highlighting potential research directions that can advance textual and multimodal QA systems in the context of adversarial challenges.

Artificial neural networks (ANNs) are inspired by human learning. However, unlike human education, classical ANN does not use a curriculum. Curriculum learning (CL) refers to the process of ANN training in which samples are used in a meaningful order. When using CL, training begins with a subset of the dataset and new samples are added throughout the training, or training begins with the entire dataset and the number of samples used is reduced. With these changes in training dataset size, better results can be obtained with curriculum, anti-curriculum, or random-curriculum methods than the vanilla method. However, a generally efficient CL method for various architectures and datasets is not found. In this article, we propose cyclical CL (CCL), in which the data size used during training changes cyclically rather than simply increasing or decreasing. Instead of using only the vanilla method or only the curriculum method, using both methods cyclically like in CCL provides more successful results. We tested the method on 18 different datasets and 15 architectures in image and text classification tasks and obtained more successful results than no-CL and existing CL methods. We also have shown theoretically that it is less erroneous to apply CL and vanilla cyclically instead of using only CL or only the vanilla method. The code of the cyclical curriculum is available at https://github.com/CyclicalCurriculum/Cyclical-Curriculum.

Math Word Problem (MWP) is a challenging Natural Language Processing (NLP) task. Existing MWP solvers have shown that current models need to generalize better and obtain higher performances. In this study, we aim to enrich existing MWP datasets with high-quality data, which may improve MWP solvers’ performances. We propose several data augmentation methods by applying minor modifications to the problem texts and equations of English MWPs datasets which contain equations with one unknown. Extensive experiments on two MWPs datasets have shown that data created by augmented methods have considerably improved performance. Moreover, further increasing the training samples by combining the samples generated by the proposed augmentation methods provides further performance improvements.

Appropriate reviewer assignment significantly impacts the quality of proposal evaluation, as accurate and fair reviews are contingent on their assignment to relevant reviewers. The crucial task of assigning reviewers to submitted proposals is the starting point of the review process and is also known as the reviewer assignment problem (RAP). Due to the obvious restrictions of manual assignment, journal editors, conference organizers, and grant managers demand automatic reviewer assignment approaches. Many studies have proposed assignment solutions in response to the demand for automated procedures since 1992. The primary objective of this survey paper is to provide scholars and practitioners with a comprehensive overview of available research on the RAP. To achieve this goal, this article presents an in-depth systematic review of 103 publications in the field of reviewer assignment published in the past three decades and available in the Web of Science, Scopus, ScienceDirect, Google Scholar, and Semantic Scholar databases. This review paper classified and discussed the RAP approaches into two broad categories and numerous subcategories based on their underlying techniques. Furthermore, potential future research directions for each category are presented. This survey shows that the research on the RAP is becoming more significant and that more effort is required to develop new approaches and a framework.

In this paper, we introduce a Turkish adaption of CLIP (Contrastive Language-Image Pre-Training). Our approach is to train a model with the same output space as the Text encoder of the CLIP model while processing Turkish input. For this, we collected 2.5M unique English-Turkish data. The model we named TrCLIP performed 71% in CIFAR100, 86% in VOC2007, and 47% in FER2013 as zero-shot accuracy. We have examined its performance on e-commerce data and a vast domain-independent dataset in image and text retrieval tasks. The model can work in Turkish without any extra fine-tuning. Models and dataset can be reachable from https://github.com/yusufani/TrCLIP.

Abstract In this study, two GRU variants named GRU1 and GRU2 are proposed by employing simple changes to the internal structure of the standard GRU, which is one of the popular RNN variants. Comparative experiments are conducted on four problems: language modeling, question answering, addition task, and sentiment analysis. Moreover, in the addition task, curriculum learning and anti‐curriculum learning strategies, which extend the training data having examples from easy to hard or from hard to easy, are comparatively evaluated. Accordingly, the GRU1 and GRU2 variants outperformed the standard GRU. In addition, the curriculum learning approach, in which the training data is expanded from easy to difficult, improves the performance considerably.

Abstract Being similar to and inspired from the process of human education, curriculum learning methods ‐or CL methods‐sort the input examples from easy to difficult, then add them to the training set in that order. Considering the fact that CL research is most concerned with determining the direction (from easy to difficult vs. from difficult to easy) and the criteria of this sorting, vast and various studies have emerged in the literature addressing both types of sorting. However, this results in a contradiction that demands finding a common aspect of ordering in both directions. This study argues that this required common aspect lies in the gradual enlargement of training. In other words, it is claimed that the success of CL methods does not depend on which criteria or in which direction the ordering is made. Extensive experiments have been conducted on various datasets using different deep learning models in order to test this claim. It was observed that random ordering had achieved competitive results with CL methods. Moreover, random ordering proved to be faster than other CL methods as it eliminates the cost of sorting computation. Based on these results, using random ordered growing sets as a baseline in future CL studies is recommended. Moreover, the possibly to improve the optimization performance via training with growing sets in theoretical perspective is also explained.

Paraphrase generation is an important NLP task of generation a sentence that has the same meaning as a given sentence. But, there are few studies on Turkish in this field. In this study, we present the most comprehensive study on Turkish. We focused on question sentences due to the resources we could access. First of all, since there is not enough large parallel paraphrase dataset for Turkish, we created several datasets with various methods. We compared the success of these datasets using the BERT2BERT architecture. We found the combination of automatically generated and manually generated datasets to be the most successful dataset. Using this dataset, we compared the success of the most popular architectures in the field. We found that MBart is the most successful model according to BLEU/Rouge criteria and BERT2BERT according to human evaluation.

Recurrent Neural Network (RNN) is a widely used deep learning architecture applied to sequence learning problems. However, it is recognized that RNNs suffer from exploding and vanishing gradient problems that prohibit the early layers of the network from learning the gradient information. GRU networks are particular kinds of recurrent networks that reduce the short-comings of these problems. In this study, we propose two variants of the standard GRU with simple but effective modifications. We applied an empirical approach and tried to determine the effectiveness of the current units and recurrent units of gates by giving different coefficients. Interestingly, we realize that applying such minor and simple changes to the standard GRU provides notable improvements. We comparatively evaluate the standard GRU with the proposed two variants on four different tasks: (1) sentiment classification on the IMDB movie review dataset, (2) language modeling task on Penn TreeBank (PTB) dataset, (3) sequence to sequence addition problem, and (4) question answering problem on Facebook’s bAbitasks dataset. The evaluation results indicate that the proposed two variants of GRU consistently outperform standard GRU.

The performance of the ensemble algorithms is related with the individual accuracy of the base learners and their results diversity. Individual accuracy of a base learner is directly related to the similarity between the original training set and the base learner's training set. When a modified training set by randomly selecting features/classes/samples is given to the base learners, the diversity is created but the individual accuracy is decreased. From this point of view, different ensemble algorithms can be seen as a selection between having more accurate but less diverse base learners and having more diverse but less accurate base learners. We propose a meta ensemble method named as improved space forest which adds generated and (hopefully) more accurate features to the original features. The new features are obtained from randomly selected original features. When the new features are more distinctive than the original ones, they are selected by the learners. So, the ensemble may have more accurate base learners. However, a different improved space is generated for each learner to create diversity. The proposed method can be used with different ensemble methods. We compared original and improved space versions of bagging, random forest, and rotation forest algorithms. Improved space versions have generally better or comparable results than the original ones. We also present a theoretical framework to analyze the individual accuracies and diversities of the base learners.

Detection of thermal targets for search and rescue robots is very important to be able to save more lives. Because the living human body is at a certain temperature, each thermal target point implies possible victim. Robots produced for search and rescue are expected to be able to perceive and steer toward the thermal targets. The focus of this work, which is also a criterion of RoboCup competitions, is the development of an exploration method for the determination of thermal targets. An algorithm has been developed which relies on giving travel priority to the thermal information emitting targets in the environment. So that the victims can be detected more effectively. Additional methodsg, such as human detection from image processing, detecting carbon dioxide gas, motion detection, etc., can be used to identify the victim, in consideration of the fact that every thermal target in the environment may not be human bein This study only involves detecting thermal targets and directing the mobile robots to them. Successful results are ensured by making the method more stable thanks to tests in both the real environment and the simulation environment. Gazebo is used as the simulation environment, and a differential drive mobile robot with a thermal camera is used for real environment experiments. Since there is no thermal camera in Gazebo simulation environment, a system was designed to represent thermal targets. This system is based on obtaining representative thermal images by applying various filters to normal camera images.

Urban search and rescue robots explore the area which they don't know. They must localize themselves, map the environment, and choose their targets. The usage of robot teams can be very effective for large areas instead of a single robot. Robot teams can be managed with distributed or centric methodologies. In a distributed architecture, each robot should be self-sufficient by means of all search tasks. This also means that each robot needs a rich computational power. Moreover, an optimal exploration strategy requires communication between all the robots. A common way to get such an architecture is applying centric approaches. In centric approaches, each robot can be seen as a mobile sensor with little computational power. They send their measures to the center. The map is generated at the center. Navigation commands are generated at the center according to the exploration strategy. ROS is a very common platform for robotic researchers. It includes several single robot mapping algorithms. But, there is no common mapping algorithm for centric approaches. In this study, we developed a multi-robot version of Hector mapping which is widely used in most robotic researches. For the real-time running ability, we parallelized its optimization procedure. The experimental results shows the effectiveness of our proposed architecture.

Exploration is defined as the selection of target points that yield the biggest contribution to a specific gain function at an initially unknown environment. Exploration for autonomous mobile robots is closely related to mapping, navigation, localization and obstacle avoidance. In this study an autonomous frontier-based exploration strategy is implemented. Frontiers are defined as the border points that are calculated throughout the mapping and navigation stage between known and unknown areas. Frontier-based exploration implementation is compatible with the Robot Operating System (ROS). Also in this study, real robot platform is utilized for testing and the effect of different frontier target assignment approaches are comparatively analyzed by means of total path length and thereby total exploration time.

The extended space forest is a new method for decision tree construction in which training is done with input vectors including all the original features and their random combinations. The combinations are generated with a difference operator applied to random pairs of original features. The experimental results show that extended space versions of ensemble algorithms have better performance than the original ensemble algorithms. To investigate the success dynamics of the extended space forest, the individual accuracy and diversity creation powers of ensemble algorithms are compared. The Extended Space Forest creates more diversity when it uses all the input features than Bagging and Rotation Forest. It also results in more individual accuracy when it uses random selection of the features than Random Subspace and Random Forest methods. It needs more training time because of using more features than the original algorithms. But its testing time is lower than the others because it generates less complex base learners.

A new family of algorithm called Cline that provides a number of methods to construct and use multivariate decision trees is presented. We report experimental results for two types of data: synthetic data to visualize the behavior of the algorithms and publicly available eight data sets. The new methods have been tested against 23 other decision-tree construction algorithms based on benchmark data sets. Empirical results indicate that our approach achieves better classification accuracy compared to other algorithms.