Hard Datasets Research Articles

Switching Self-Attention Text Classification Model with Innovative Reverse Positional Encoding for Right-to-Left Languages: A Focus on Arabic Dialects

Transformer models have emerged as frontrunners in the field of natural language processing, primarily due to their adept use of self-attention mechanisms to grasp the semantic linkages between words in sequences. Despite their strengths, these models often face challenges in single-task learning scenarios, particularly when it comes to delivering top-notch performance and crafting strong latent feature representations. This challenge is more pronounced in the context of smaller datasets and is particularly acute for under-resourced languages such as Arabic. In light of these challenges, this study introduces a novel methodology for text classification of Arabic texts. This method harnesses the newly developed Reverse Positional Encoding (RPE) technique. It adopts an inductive-transfer learning (ITL) framework combined with a switching self-attention shared encoder, thereby increasing the model’s adaptability and improving its sentence representation accuracy. The integration of Mixture of Experts (MoE) and RPE techniques empowers the model to process longer sequences more effectively. This enhancement is notably beneficial for Arabic text classification, adeptly supporting both the intricate five-point and the simpler ternary classification tasks. The empirical evidence points to its outstanding performance, achieving accuracy rates of 87.20% for the HARD dataset, 72.17% for the BRAD dataset, and 86.89% for the LABR dataset, as evidenced by the assessments conducted on these datasets.

Evaluation of Different Stemming Techniques on Arabic Customer Reviews

Customer opinion and reviews play a vital role in marketing expansion. Big companies all over the world assign a lot of their efforts to analyzing customers’ feedback to keep track of their needs. Natural Language Processing (NLP) is widely used to analyze such review texts. Arabic customer analysis and classification also began to gain researchers’ attention due to the wide range of Arabic language speakers. Working with Arabic Language is a very challenging task because of the orthographic nature of Arabic. Also, customers often write their reviews in their dialectical style, which often diverts from standard Arabic. This study presents a method to classify Arabic customer reviews using four classifiers (K-nearest Neighbor (KNN), Support Vector Machine (SVM), Logistic Regression (RL), and Naïve Bayes (NB)). The classification is implemented with three stemming techniques (Snowball, Khoja, and Tashaphyne). The HARD dataset is adopted to perform the experiments. The results stated that the stemming methods can enhance classification performance despite the complexity of Arabic scripts and dialects. This work sheds light on utilizing and investigating more machine learning (ML) techniques and evaluating the results.

Switch-Transformer Sentiment Analysis Model for Arabic Dialects That Utilizes a Mixture of Experts Mechanism

In recent years, models such as the transformer have demonstrated impressive capabilities in the realm of natural language processing. However, these models are known for their complexity and the substantial training they require. Furthermore, the self-attention mechanism within the transformer, designed to capture semantic relationships among words in sequences, faces challenges when dealing with short sequences. This limitation hinders its effectiveness in five-polarity Arabic sentiment analysis (SA) tasks. The switch-transformer model has surfaced as a potential substitute. Nevertheless, when employing one-task learning for their training, these models frequently face challenges in presenting exceptional performances and encounter issues when producing resilient latent feature representations, particularly in the context of small-size datasets. This challenge is particularly prominent in the case of the Arabic dialect, which is recognized as a low-resource language. In response to these constraints, this research introduces a novel method for the sentiment analysis of Arabic text. This approach leverages multi-task learning (MTL) in combination with the switch-transformer shared encoder to enhance model adaptability and refine sentence representations. By integrating a mixture of experts (MoE) technique that breaks down the problem into smaller, more manageable sub-problems, the model becomes skilled in managing extended sequences and intricate input–output relationships, thereby benefiting both five-point and three-polarity Arabic sentiment analysis tasks. The proposed model effectively identifies sentiment in Arabic dialect sentences. The empirical results underscore its exceptional performance, with accuracy rates reaching 84.02% for the HARD dataset, 67.89% for the BRAD dataset, and 83.91% for the LABR dataset, as demonstrated by the evaluations conducted on these datasets.

Enhancing Arabic E-Commerce Review Sentiment Analysis Using a hybrid Deep Learning Model and FastText word embedding

The usage of NLP is shown in sentiment analysis (SA). SA extracts textual views. Arabic SA is challenging because of ambiguity, dialects, morphological variation, and the need for more resources available. The application of convolutional neural networks to Arabic SA has shown to be successful. Hybrid models improve single deep learning models. By layering many deep learning ensembles, earlier deep learning models should achieve higher accuracy. This research successfully predicted Arabic sentiment using CNN, LSTM, GRU, BiGRU, BiLSTM, CNN-BiGRU, CNN-GRU, CNN-LSTM, and CNN-biLSTM. Two enormous datasets, including the HARD and BRAD datasets, are used to evaluate the effectiveness of the proposed model. The findings demonstrated that the provided model could interpret the feelings conveyed in Arabic. The proposed procedure kicks off with the extraction of Arabert model features. After that, we developed and trained nine deep-learning models, including CNN, LSTM, GRU, BiGRU, BiLSTM, CNN-BiGRU, CNN-GRU, CNN-LSTM, and CNN-biLSTM. Concatenating the FastText and GLOVE as word embedding models. By a margin of 0.9112, our technique surpassed both standard forms of deep learning.

A dependency-based hybrid deep learning framework for target-dependent sentiment classification

One of the main challenges in target-dependent sentiment classification (TDSC) is dealing with sentences that contain multiple targets with varying polarities. Traditional sentiment analysis has shown the effectiveness of language characteristics. Therefore, we propose a method to extract target semantic-related tokens from sentences in order to simplify the sentiment classification task. To achieve this, we establish six grammatical principles that utilize grammatical knowledge to filter the relevant descriptions of targets. Since a target is typically a noun and acts as a subject, we summarize the six rules to extract the contexts contained in the objects and subordinate clauses. We use dependency parsing to analyze the grammatical relations between the target and its context. We design a data pre-processing method called Text Filtering (TF) to automate this procedure. After executing the TF algorithm, we pass the target-related words to a simple classifier to predict their sentiment polarities. Rather than feeding these features directly to a network and letting it learn features on its own, our approach employs dependency relations to extract context linked to the target. This provides the network with meaningful and representative features, resulting in superior results. We conduct ablation studies to investigate the effectiveness of the proposed TF algorithm. In the restaurant hard dataset, our approach improves accuracy by 13.76% and macro-F1 by 14.65% compared to a CNN-based method where TF is not implemented.

Cell-penetrating peptides predictors: A comparative analysis of methods and datasets.

Cell-Penetrating Peptides (CPP) are emerging as an alternative to small-molecule drugs to expand the range of biomolecules that can be targeted for therapeutic purposes. Due to the importance of identifying and designing new CPP, a great variety of predictors have been developed to achieve these goals. To establish a ranking for these predictors, a couple of recent studies compared their performances on specific datasets, yet their conclusions cannot determine if the ranking obtained is due to the model, the set of descriptors or the datasets used to test the predictors. We present a systematic study of the influence of the peptide sequence's similarity of the datasets on the predictors' performance. The analysis reveals that the datasets used for training have a stronger influence on the predictors performance than the model or descriptors employed. We show that datasets with low sequence similarity between the positive and negative examples can be easily separated, and the tested classifiers showed good performance on them. On the other hand, a dataset with high sequence similarity between CPP and non-CPP will be a hard dataset, and it should be the one to be used for assessing the performance of new predictors.

User Opinion Prediction for Arabic Hotel Reviews Using Lexicons and Artificial Intelligence Techniques

Opinion mining refers to the process that helps to identify and to classify users’ emotions and opinions from any source, such as an online review. Thus, opinion mining provides organizations with an insight into their reputation based on previous customers’ opinions regarding their services or products. Automating opinion mining in different languages is still an important topic of interest for scientists, including those using the Arabic language, especially since potential customers mostly do not rate their opinion explicitly. This study proposes an ensemble-based deep learning approach using fastText embeddings and the proposed Arabic emoji and emoticon opinion lexicon to predict user opinion. For testing purposes, the study uses the publicly available Arabic HARD dataset, which includes hotel reviews associated with ratings, starting from one to five. Then, by employing multiple Arabic resources, it experiments with different generated features from the HARD dataset by combining shallow learning with the proposed approach. To the best of our knowledge, this study is the first to create a lexicon that considers emojis and emoticons for its user opinion prediction. Therefore, it is mainly a helpful contribution to the literature related to opinion mining and emojis and emoticons lexicons. Compared to other studies found in the literature related to the five-star rating prediction using the HARD dataset, the accuracy of the prediction using the proposed approach reached an increase of 3.21% using the balanced HARD dataset and an increase of 2.17% using the unbalanced HARD dataset. The proposed work can support a new direction for automating the unrated Arabic opinions in social media, based on five rating levels, to provide potential stakeholders with a precise idea about a service or product quality, instead of spending much time reading other opinions to learn that information.

Using AraGPT and ensemble deep learning model for sentiment analysis on Arabic imbalanced dataset

With the fast growth of mobile technology, social media has become important for people to share their thoughts and feelings. Businesses and governments can make better strategic decisions when they know what the public thinks. Because of this, sentiment analysis is an important tool for figuring out how different people’s opinions are. This article presents a deeplearning ensemble model for sentiment analysis. The ensemble model proposed consists of three deep-learning models, Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM), as base classifiers. AraBERT is responsible for presenting the textual input data into representative embeddings. The stacking ensemble model then captures the long-range dependencies in the embedding for a given class. As a meta-classifier, Support Vector Machine (SVM) then combines the predictions made by the stacking deep learning model. In addition, data augmentation with AraGPT was implemented to address the imbalanced dataset issues. The experimental results demonstrate that the proposed model outperforms the state-of-the-art models with an accuracy of 88.89%, 90.88%, and 88.23% on the HARD, BRAD, and Twitter datasets, respectively.

Multitasking Learning Model Based on Hierarchical Attention Network for Arabic Sentiment Analysis Classification

Limited approaches have been applied to Arabic sentiment analysis for a five-point classification problem. These approaches are based on single task learning with a handcrafted feature, which does not provide robust sentence representation. Recently, hierarchical attention networks have performed outstandingly well. However, when training such models as single-task learning, these models do not exhibit superior performance and robust latent feature representation in the case of a small amount of data, specifically on the Arabic language, which is considered a low-resource language. Moreover, these models are based on single task learning and do not consider the related tasks, such as ternary and binary tasks (cross-task transfer). Centered on these shortcomings, we regard five ternary tasks as relative. We propose a multitask learning model based on hierarchical attention network (MTLHAN) to learn the best sentence representation and model generalization, with shared word encoder and attention network across both tasks, by training three-polarity and five-polarity Arabic sentiment analysis tasks alternately and jointly. Experimental results showed outstanding performance of the proposed model, with high accuracy of 83.98%, 87.68%, and 84.59 on LABR, HARD, and BRAD datasets, respectively, and a minimum macro mean absolute error of 0.632% on the Arabic tweets dataset for five-point Arabic sentiment classification problem.

ACWFace: efficient and lightweight face detector based on RetinaFace

Lightweight face detection algorithms that typically utilize convolutional neural network to find out all faces from the entire vision range. However, compared with more accurate and heavy algorithms, the performance of existing lightweight networks is still left far behind. Toward this end, we propose a lightweight and efficient single-stage face detector, named ACWFace, which explores the effects of attention, context module, and weighted feature fusion based on RetinaFace. First, efficient dual attention module is designed to further explore the potential of channel attention and spatial attention by introducing adaptive convolution kernel. Second, extended context module and shuffled context module are proposed to enlarge the receptive field and increase the information intersection between branches. Finally, weighted-fusion feature pyramid network is utilized to solve the features fusion of different scales equally by introducing the feature fusion module. Experiments on the easy, medium, and hard datasets of WIDER FACE validation partition show that our ACWFace outperforms RetinaFace average precision by 1.0%, 1.1%, and 1.8% while it achieves a great growth of 0.6%, 6.5%, and 3.0% on annotated faces in the wild, PASCAL face, and face detection data set and benchmark datasets, respectively.

Three-Filters-to-Normal: An Accurate and Ultrafast Surface Normal Estimator

This paper proposes three-filters-to-normal (3F2N), an accurate and ultrafast surface normal estimator (SNE), which is designed for structured range sensor data, e.g., depth/disparity images. 3F2N SNE computes surface normals by simply performing three filtering operations (two image gradient filters in horizontal and vertical directions, respectively, and a mean/median filter) on an inverse depth image or a disparity image. Despite the simplicity of 3F2N SNE, no similar method already exists in the literature. To evaluate the performance of our proposed SNE, we created three large-scale synthetic datasets (easy, medium and hard) using 24 3D mesh models, each of which is used to generate 1800--2500 pairs of depth images (resolution: 480X640 pixels) and the corresponding ground-truth surface normal maps from different views. 3F2N SNE demonstrates the state-of-the-art performance, outperforming all other existing geometry-based SNEs, where the average angular errors with respect to the easy, medium and hard datasets are 1.66 degrees, 5.69 degrees and 15.31 degrees, respectively. Furthermore, our C++ and CUDA implementations achieve a processing speed of over 260 Hz and 21 kHz, respectively. Our datasets and source code are publicly available at sites.google.com/view/3f2n.

A competitive optimization approach for data clustering and orthogonal non-negative matrix factorization

Partitioning a given data-set into subsets based on similarity among the data is called clustering. Clustering is a major task in data mining and machine learning having many applications such as text retrieval, pattern recognition, and web mining. Here, we briefly review some clustering related problems (k-means, normalized k-cut, orthogonal non-negative matrix factorization, ONMF, and isoperimetry) and describe their connections. We formulate the relaxed mean version of the isoperimetry problem as an optimization problem with non-negative orthogonal constraints. We first make use of a gradient-based optimization algorithm to solve this kind of a problem, and then apply a post-processing technique to extract a solution of the clustering problem. Also, we propose a simplified approach to improve upon solution of the 2-dimensional clustering problem, using the N-nearest neighbor graph. Inspired by this technique, we apply a multilevel method for clustering a given data-set to reduce the size of the problem by grouping a number of similar vertices. The number is determined based on two values, namely, the maximum and the average of the edge weights of the vertices connected to a selected vertex. In addition, using the connections between ONMF and k-means and between k-means and the isoperimetry problem, we propose an algorithm to solve the ONMF problem. A comparative performance analysis of our approach with other related methods shows outperformance of our approach, in terms of the obtained misclassification error rate and Rand index, on both benchmark and randomly generated problems as well as hard synthetic data-sets.

Context-aware MDNet for target tracking in UAV remote sensing videos

ABSTRACTWith the maturity and popularity of unmanned aerial vehicle (UAV) technology, remote sensing target tracking in the aerial videos from UAVs has drawn much attention in recent years. Given that UAV aerial video has low resolution, multiple similar disruptors and rapid perspective changes, and the main tracking methods in this research field generally have low tracking performance and timeliness, we propose a remote sensing target tracking method for UAV aerial video based on a context-aware multi-domain convolutional neural network (CAMD). First, in the design of the tracking network structure, we fuse multiple convolutional layers using residual connections to boost the effectiveness of regression learning. Then, in the offline pretraining stage, a Rotation Adversarial Autoencoders (RAAE) was adopted to generate typical easily confused negative samples for enhancing the capacity to distinguish between targets and the background interference. Finally, when the estimated target score was in the ‘fuzzy interval’, we introduced a response-adaptive context-aware correlation filter (RA-CACF) module into our network architecture to improve the tracking performance. Qualitative and quantitative evaluations using public and homemade hard datasets demonstrate that the proposed method can achieve high accuracy and efficiency results compared to state-of-the-art trackers.

A Better Method of Applying OLS to the CAPM, Prediction, and Forecasting

This paper formulates a weighting function from conventional least squares (LS) and combines it with estimation theory to provide the statistical estimate, expected value, and variance of any point on the polynomial constructed for fitting a set of existing data. This solves the problem of the missing variance at arbitrary points on the polynomial from LS derived by Gauss. The method includes three LS tricks: (a) Reframing LS from creating a polynomial for fitting existing data to estimating an already existing polynomial corrupted with statistically described sampling errors. (b) Restructuring LS processing from approximating polynomial coefficients to creating a weighting function for estimating the independent variable at any point on the LS polynomial. (c) Averaging the statistical deviations from the estimated LS polynomial to estimate the variance of sampling errors. The method is based on two cold hard CAPM data sets: (a) samples of the asset corrupted by statistically described errors and (b) deterministic samples of the corresponding market. Hand-waving arguments about market forces and investor behavior apply only after LS processing, not before or during. An example of the technique applied to fictitious sales as a function of GNP shows the technique applies to virtually any problem addressed by polynomial LS.

Remote Sensing Target Tracking in UAV Aerial Video Based on Saliency Enhanced MDnet

Remote sensing target tracking in the aerial video from unmanned aerial vehicles (UAV) plays a key role in public security. As the UAV aerial video has rapid changes in scale and perspective, few pixels in the target region, and multiple similar disruptors, and the main tracking methods in this research field generally have relatively low tracking performance and timeliness, we propose a remote sensing target tracking method for the UAV aerial video based on a saliency enhanced multi-domain convolutional neural network (SEMD). First, in the pre-training stage, we combine the least squares generative adversarial networks (LSGANs) with a multi-orientation Gaussian Pyramid to augment typical easily confused negative samples for enhancing the capacity to distinguish between targets and the background. Then, a saliency module was integrated into our tracking network architecture to boost the saliency of the feature map, which can improve the representation power of a rapid dynamic change target. Finally, in the stage for generating tracking samples, we implemented a local weight allocation model to screen for hard negative samples. This approach can not only improve the stability in tracking but also boost efficiency. The comprehensive evaluations of public and homemade hard datasets demonstrate that the proposed method can achieve high accuracy and efficiency results compared with state-of-the-art methods.

Experimental identification of hard data sets for classification and feature selection methods with insights on method selection

The paper reports an experimentally identified list of benchmark data sets that are hard for representative classification and feature selection methods. This was done after systematically evaluating a total of 48 combinations of methods, involving eight state-of-the-art classification algorithms and six commonly used feature selection methods, on 129 data sets from the UCI repository (some data sets with known high classification accuracy were excluded). In this paper, a data set for classification is called hard if none of the 48 combinations can achieve an AUC over 0.8 and none of them can achieve an F-Measure value over 0.8; it is called easy otherwise. A total of 15 out of the 129 data sets were found to be hard in that sense. This paper also compares the performance of different methods, and it produces rankings of classification methods, separately on the hard data sets and on the easy data sets. This paper is the first to rank methods separately for hard data sets and for easy data sets. It turns out that the classifier rankings resulting from our experiments are somehow different from those in the literature and hence they offer new insights on method selection. It should be noted that the Random Forest method remains to be the best in all groups of experiments.

Kernel Oriented Multivariate Feature Selection for Breast Cancer Data Classification via MRMR Filter

A feature selection technique is highly preferred preceding data classification to improve prediction performance especially in the high dimensional space. In general, filter techniques can be considered as essential or assistant selection system on account of their effortlessness, adaptability, and low computational many-sided quality. Nonetheless, a progression of inconsequential cases demonstrates that filter techniques result in less precise execution since they disregard the conditions of features. Albeit few publications have committed their regard for uncover the relationship of features by multivariate-based techniques, these strategies depict connections among elements just by linear techniques. While straightforward linear combination relationship limits the transformation in execution. In this paper, we utilized kernel method for svm-RFE with MRMR way to deal with find inalienable nonlinear connections among features and also amongst feature and target. So as to uncover the viability of our technique we played out a few analyses and thought about the outcomes between our technique and other aggressive multivariatebased features selectors. In our examination, we utilized three classifiers (support vector machine, neural system and average perceptron) on two gathering datasets, to be specific two-class and multi-class datasets (principally focused on svm). Exploratory results show that the execution of our technique is superior to anything others, particularly on three hard group datasets, to be specific Wang’s Breast Cancer, Gordon’s Lung Adenocarcinoma and Pomeroy’s Medulloblastoma. Note: Entire Implementation was developed with MS Machine learning studio.

ReigSAC: fast discrimination of spurious keypoint correspondences on planar surfaces

Various methods were proposed to detect/match special interest points (keypoints) in images and some of them (e.g., SIFT and SURF) are among the most cited techniques in computer vision research. This paper describes an algorithm to discriminate between genuine and spurious keypoint correspondences on planar surfaces. We draw random samples of the set of correspondences, from which homographies are obtained and their principal eigenvectors extracted. Density estimation on that feature space determines the most likely true transform. Such homography feeds a cost function that gives the goodness of each keypoint correspondence. Being similar to the well-known RANSAC strategy, the key finding is that the main eigenvector of the most (genuine) homographies tends to represent a similar direction. Hence, density estimation in the eigenspace dramatically reduces the number of transforms actually evaluated to obtain reliable estimations. Our experiments were performed on hard image data sets, and pointed that the proposed approach yields effectiveness similar to the RANSAC strategy, at significantly lower computational burden, in terms of the proportion between the number of homographies generated and those that are actually evaluated.

Cognitive Ability-Demand Gap Analysis With Latent Response Models

A better understanding of human cognitive ability-demand gap (ADG) is critical in designing assistive technology solution that is accurate and adaptive over a wide range of human-agent interaction. The main goal is to design systems that can adapt with the user's abilities and needs over a range of cognitive tasks. It will also enable the system to provide feedback consistent with the situation. However, the latent structure and relationship between human ability to respond to cognitive task (demand on human by the agent) remains unknown. Robust modeling of cognitive ADG will be a paradigm shift from the current trends in assistive technology design. The key idea is to estimate the gap, based on human-agent cognitive task interaction. In particular, latent response model was adopted to quantify the gap. First, we used one parameter Rasch model and extended Rasch model (rating scale model, partial credit model) with dichotomous and polytomous responses, respectively. Residues between expected and observed ability scores were considered as gap parameter in case of dichotomous response. In extended Rasch modeling, response latitudes are considered as an indicator of the gap. Additionally, we performed model fitting, standard error measurement, kernel density estimation, and differential item functioning to test the suitability of Rasch model. Empirical analyses on a number of data set show that proposed analytical method can model the cognitive ADG from dichotomous and polytomous responses. In dichotomous case, the model better fits for mixed responses (combination of easy, medium, and hard) data set rather than monotonic (e.g., only easy) data. Results show that Rasch model can be reliably used to estimate cognitive gap with different cognitive task types.

Modeling and mining term association for improving biomedical information retrieval performance

BackgroundThe growth of the biomedical information requires most information retrieval systems to provide short and specific answers in response to complex user queries. Semantic information in the form of free text that is structured in a way makes it straightforward for humans to read but more difficult for computers to interpret automatically and search efficiently. One of the reasons is that most traditional information retrieval models assume terms are conditionally independent given a document/passage. Therefore, we are motivated to consider term associations within different contexts to help the models understand semantic information and use it for improving biomedical information retrieval performance.ResultsWe propose a term association approach to discover term associations among the keywords from a query. The experiments are conducted on the TREC 2004-2007 Genomics data sets and the TREC 2004 HARD data set. The proposed approach is promising and achieves superiority over the baselines and the GSP results. The parameter settings and different indices are investigated that the sentence-based index produces the best results in terms of the document-level, the word-based index for the best results in terms of the passage-level and the paragraph-based index for the best results in terms of the passage2-level. Furthermore, the best term association results always come from the best baseline. The tuning number k in the proposed recursive re-ranking algorithm is discussed and locally optimized to be 10.ConclusionsFirst, modelling term association for improving biomedical information retrieval using factor analysis, is one of the major contributions in our work. Second, the experiments confirm that term association considering co-occurrence and dependency among the keywords can produce better results than the baselines treating the keywords independently. Third, the baselines are re-ranked according to the importance and reliance of latent factors behind term associations. These latent factors are decided by the proposed model and their term appearances in the first round retrieved passages.