Search task success evaluation by exploiting multi-view active semi-supervised learning

Search task success rate is an important indicator to measure the performance of search engines. In contrast to most of the previous approaches that rely on labeled search tasks provided by users or third-party editors, this paper attempts to improve the performance of search task success evaluation by exploiting unlabeled search tasks that are existing in search logs as well as a small amount of labeled ones. Concretely, the Multi-view Active Semi-Supervised Search task Success Evaluation (MA4SE) approach is proposed, which exploits labeled data and unlabeled data by integrating the advantages of both semi-supervised learning and active learning with the multi-view mechanism. In the semi-supervised learning part of MA4SE, we employ a multi-view semi-supervised learning approach that utilizes different parameter configurations to achieve the disagreement between base classifiers. The base classifiers are trained separately from the pre-defined action and time views. In the active learning part of MA4SE, each classifier received from semi-supervised learning is applied to unlabeled search tasks, and the search tasks that need to be manually annotated are selected based on both the degree of disagreement between base classifiers and a regional density measurement. We evaluate the proposed approach on open datasets with two different definitions of search tasks success. The experimental results show that MA4SE outperforms the state-of-the-art semi-supervised search task success evaluation approach.     

Text classification from unlabeled documents with bootstrapping and feature projection techniques

Many machine learning algorithms have been applied to text classification tasks. In the machine learning paradigm, a general inductive process automatically builds a text classifier by learning, generally known as supervised learning. However, the supervised learning approaches have some problems. The most notable problem is that they require a large number of labeled training documents for accurate learning. While unlabeled documents are easily collected and plentiful, labeled documents are difficultly generated because a labeling task must be done by human developers. In this paper, we propose a new text classification method based on unsupervised or semi-supervised learning. The proposed method launches text classification tasks with only unlabeled documents and the title word of each category for learning, and then it automatically learns text classifier by using bootstrapping and feature projection techniques. The results of experiments showed that the proposed method achieved reasonably useful performance compared to a supervised method. If the proposed method is used in a text classification task, building text classification systems will become significantly faster and less expensive.     

AdaSL: An Unsupervised Domain Adaptation framework for Arabic multi-dialectal Sequence Labeling

Dialectal Arabic (DA) refers to varieties of everyday spoken languages in the Arab world. These dialects differ according to the country and region of the speaker, and their textual content is constantly growing with the rise of social media networks and web blogs. Although research on Natural Language Processing (NLP) on standard Arabic, namely Modern Standard Arabic (MSA), has witnessed remarkable progress, research efforts on DA are rather limited. This is due to numerous challenges, such as the scarcity of labeled data as well as the nature and structure of DA. While some recent works have reached decent results on several DA sentence classification tasks, other complex tasks, such as sequence labeling, still suffer from weak performances when it comes to DA varieties with either a limited amount of labeled data or unlabeled data only. Besides, it has been shown that zero-shot transfer learning from models trained on MSA does not perform well on DA. In this paper, we introduce AdaSL, a new unsupervised domain adaptation framework for Arabic multi-dialectal sequence labeling, leveraging unlabeled DA data, labeled MSA data, and existing multilingual and Arabic Pre-trained Language Models (PLMs). The proposed framework relies on four key components: (1) domain adaptive fine-tuning of multilingual/MSA language models on unlabeled DA data, (2) sub-word embedding pooling, (3) iterative self-training on unlabeled DA data, and (4) iterative DA and MSA distribution alignment. We evaluate our framework on multi-dialectal Named Entity Recognition (NER) and Part-of-Speech (POS) tagging tasks.The overall results show that the zero-shot transfer learning, using our proposed framework, boosts the performance of the multilingual PLMs by 40.87% in macro-F1 score for the NER task, while it boosts the accuracy by 6.95% for the POS tagging task. For the Arabic PLMs, our proposed framework increases performance by 16.18% macro-F1 for the NER task and 2.22% accuracy for the POS tagging task, and thus, achieving new state-of-the-art zero-shot transfer learning performance for Arabic multi-dialectal sequence labeling.     

Unsupervised graph-based rank aggregation for improved retrieval

This paper presents a robust and comprehensive graph-based rank aggregation approach, used to combine results of isolated ranker models in retrieval tasks. The method follows an unsupervised scheme, which is independent of how the isolated ranks are formulated. Our approach is able to combine arbitrary models, defined in terms of different ranking criteria, such as those based on textual, image or hybrid content representations.We reformulate the ad-hoc retrieval problem as a document retrieval based on fusion graphs, which we propose as a new unified representation model capable of merging multiple ranks and expressing inter-relationships of retrieval results automatically. By doing so, we claim that the retrieval system can benefit from learning the manifold structure of datasets, thus leading to more effective results. Another contribution is that our graph-based aggregation formulation, unlike existing approaches, allows for encapsulating contextual information encoded from multiple ranks, which can be directly used for ranking, without further computations and post-processing steps over the graphs. Based on the graphs, a novel similarity retrieval score is formulated using an efficient computation of minimum common subgraphs. Finally, another benefit over existing approaches is the absence of hyperparameters.A comprehensive experimental evaluation was conducted considering diverse well-known public datasets, composed of textual, image, and multimodal documents. Performed experiments demonstrate that our method reaches top performance, yielding better effectiveness scores than state-of-the-art baseline methods and promoting large gains over the rankers being fused, thus demonstrating the successful capability of the proposal in representing queries based on a unified graph-based model of rank fusions.     

Arabic text classification using deep learning models

Text classification or categorization is the process of automatically tagging a textual document with most relevant labels or categories. When the number of labels is restricted to one, the task becomes single-label text categorization. However, the multi-label version is challenging. For Arabic language, both tasks (especially the latter one) become more challenging in the absence of large and free Arabic rich and rational datasets. Therefore, we introduce new rich and unbiased datasets for both the single-label (SANAD) as well as the multi-label (NADiA) Arabic text categorization tasks. Both corpora are made freely available to the research community on Arabic computational linguistics. Further, we present an extensive comparison of several deep learning (DL) models for Arabic text categorization in order to evaluate the effectiveness of such models on SANAD and NADiA. A unique characteristic of our proposed work, when compared to existing ones, is that it does not require a pre-processing phase and fully based on deep learning models. Besides, we studied the impact of utilizing word2vec embedding models to improve the performance of the classification tasks. Our experimental results showed solid performance of all models on SANAD corpus with a minimum accuracy of 91.18%, achieved by convolutional-GRU, and top performance of 96.94%, achieved by attention-GRU. As for NADiA, attention-GRU achieved the highest overall accuracy of 88.68% for a maximum subsets of 10 categories on “Masrawy” dataset.     

RoSAS: Deep semi-supervised anomaly detection with contamination-resilient continuous supervision

Semi-supervised anomaly detection methods leverage a few anomaly examples to yield drastically improved performance compared to unsupervised models. However, they still suffer from two limitations: 1) unlabeled anomalies (i.e., anomaly contamination) may mislead the learning process when all the unlabeled data are employed as inliers for model training; 2) only discrete supervision information (such as binary or ordinal data labels) is exploited, which leads to suboptimal learning of anomaly scores that essentially take on a continuous distribution. Therefore, this paper proposes a novel semi-supervised anomaly detection method, which devises contamination-resilient continuous supervisory signals. Specifically, we propose a mass interpolation method to diffuse the abnormality of labeled anomalies, thereby creating new data samples labeled with continuous abnormal degrees. Meanwhile, the contaminated area can be covered by new data samples generated via combinations of data with correct labels. A feature learning-based objective is added to serve as an optimization constraint to regularize the network and further enhance the robustness w.r.t. anomaly contamination. Extensive experiments on 11 real-world datasets show that our approach significantly outperforms state-of-the-art competitors by 20%–30% in AUC-PR and obtains more robust and superior performance in settings with different anomaly contamination levels and varying numbers of labeled anomalies.     

Robust graph learning with graph convolutional network

Graph convolutional network (GCN) is a powerful tool to process the graph data and has achieved satisfactory performance in the task of node classification. In general, GCN uses a fixed graph to guide the graph convolutional operation. However, the fixed graph from the original feature space may contain noises or outliers, which may degrade the effectiveness of GCN. To address this issue, in this paper, we propose a robust graph learning convolutional network (RGLCN). Specifically, we design a robust graph learning model based on the sparse constraint and strong connectivity constraint to achieve the smoothness of the graph learning. In addition, we introduce graph learning model into GCN to explore the representative information, aiming to learning a high-quality graph for the downstream task. Experiments on citation network datasets show that the proposed RGLCN outperforms the existing comparison methods with respect to the task of node classification.     

Towards a real-time processing framework based on improved distributed recurrent neural network variants with fastText for social big data analytics

Big data generated by social media stands for a valuable source of information, which offers an excellent opportunity to mine valuable insights. Particularly, User-generated contents such as reviews, recommendations, and users’ behavior data are useful for supporting several marketing activities of many companies. Knowing what users are saying about the products they bought or the services they used through reviews in social media represents a key factor for making decisions. Sentiment analysis is one of the fundamental tasks in Natural Language Processing. Although deep learning for sentiment analysis has achieved great success and allowed several firms to analyze and extract relevant information from their textual data, but as the volume of data grows, a model that runs in a traditional environment cannot be effective, which implies the importance of efficient distributed deep learning models for social Big Data analytics. Besides, it is known that social media analysis is a complex process, which involves a set of complex tasks. Therefore, it is important to address the challenges and issues of social big data analytics and enhance the performance of deep learning techniques in terms of classification accuracy to obtain better decisions.In this paper, we propose an approach for sentiment analysis, which is devoted to adopting fastText with Recurrent neural network variants to represent textual data efficiently. Then, it employs the new representations to perform the classification task. Its main objective is to enhance the performance of well-known Recurrent Neural Network (RNN) variants in terms of classification accuracy and handle large scale data. In addition, we propose a distributed intelligent system for real-time social big data analytics. It is designed to ingest, store, process, index, and visualize the huge amount of information in real-time. The proposed system adopts distributed machine learning with our proposed method for enhancing decision-making processes. Extensive experiments conducted on two benchmark data sets demonstrate that our proposal for sentiment analysis outperforms well-known distributed recurrent neural network variants (i.e., Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Gated Recurrent Unit (GRU)). Specifically, we tested the efficiency of our approach using the three different deep learning models. The results show that our proposed approach is able to enhance the performance of the three models. The current work can provide several benefits for researchers and practitioners who want to collect, handle, analyze and visualize several sources of information in real-time. Also, it can contribute to a better understanding of public opinion and user behaviors using our proposed system with the improved variants of the most powerful distributed deep learning and machine learning algorithms. Furthermore, it is able to increase the classification accuracy of several existing works based on RNN models for sentiment analysis.     

Semi-supervised ranking aggregation

Ranking aggregation is a task of combining multiple ranking lists given by several experts or simple rankers to get a hopefully better ranking. It is applicable in several fields such as meta search and collaborative filtering. Most of the existing work is under an unsupervised framework. In these methods, the performances are usually limited especially in unreliable case since labeled information is not involved in. In this paper, we propose a semi-supervised ranking aggregation method, in which preference constraints of several item pairs are given. In our method, the aggregation function is learned based on the ordering agreement of different rankers. The ranking scores assigned by this ranking function on the labeled data should be consistent with the given pairwise order constraints while the ranking scores on the unlabeled data obey the intrinsic manifold structure of the rank items. The experimental results on toy data and the OHSUMED data are presented to illustrate the validity of our method.     

Rapid relevance classification of social media posts in disasters and emergencies: A system and evaluation featuring active,incremental and online learning

The research field of crisis informatics examines, amongst others, the potentials and barriers of social media use during disasters and emergencies. Social media allow emergency services to receive valuable information (e.g., eyewitness reports, pictures, or videos) from social media. However, the vast amount of data generated during large-scale incidents can lead to issue of information overload. Research indicates that supervised machine learning techniques are suitable for identifying relevant messages and filter out irrelevant messages, thus mitigating information overload. Still, they require a considerable amount of labeled data, clear criteria for relevance classification, a usable interface to facilitate the labeling process and a mechanism to rapidly deploy retrained classifiers. To overcome these issues, we present (1) a system for social media monitoring, analysis and relevance classification, (2) abstract and precise criteria for relevance classification in social media during disasters and emergencies, (3) the evaluation of a well-performing Random Forest algorithm for relevance classification incorporating metadata from social media into a batch learning approach (e.g., 91.28%/89.19% accuracy, 98.3%/89.6% precision and 80.4%/87.5% recall with a fast training time with feature subset selection on the European floods/BASF SE incident datasets), as well as (4) an approach and preliminary evaluation for relevance classification including active, incremental and online learning to reduce the amount of required labeled data and to correct misclassifications of the algorithm by feedback classification. Using the latter approach, we achieved a well-performing classifier based on the European floods dataset by only requiring a quarter of labeled data compared to the traditional batch learning approach. Despite a lesser effect on the BASF SE incident dataset, still a substantial improvement could be determined.     

CALIMERA: A new early time series classification method

Early time series classification is a variant of the time series classification task, in which a label must be assigned to the incoming time series as quickly as possible without necessarily screening through the whole sequence. It needs to be realized on the algorithmic level by fusing a decision-making method that detects the right moment to stop and a classifier that assigns a class label. The contribution addressed in this paper is twofold. Firstly, we present a new method for finding the best moment to perform an action (terminate/continue). Secondly, we propose a new learning scheme using classifier calibration to estimate classification accuracy. The new approach, called CALIMERA, is formalized as a cost minimization problem. Using two benchmark methodologies for early time series classification, we have shown that the proposed model achieves better results than the current state-of-the-art. Two most serious competitors of CALIMERA are ECONOMY and TEASER. The empirical comparison showed that the new method achieved a higher accuracy than TEASER for 35 out of 45 datasets and it outperformed ECONOMY in 20 out of 34 datasets.     

Measuring robustness in rank aggregation based on the error-effectiveness curve

Rank aggregation is an obligatory operation for many tasks of democratic elections, product recommendation, and gene identification. While the awareness of imperfect information in input rankings would lead to unreliable aggregation results has promoted the robust rank aggregation to become an increasingly active research topic recently. In this study, we focus on the problem of robustness measuring of rank aggregation methods. We first provide an analytical framework of rank aggregation robustness, in which we generate rankings with adjustable imperfect information and observe the response of rank aggregation methods. In particular, to quantify the robustness of rank aggregation methods, we introduce the concept of error-effectiveness curve, which presents the aggregation effectiveness under different imperfect information scenarios, and describes the ability of each method in giving stable aggregation results. By doing so, a robustness measure of rank aggregation with random error is developed. Comprehensive experimental evaluations were conducted considering synthetic datasets with various levels of random error to demonstrate the validity of the proposed measure. Exact robustness was quantified for each evaluated rank aggregation methods, and significant robustness distinctions were achieved among them.     

Spanish Corpora of tweets about COVID-19 vaccination for automatic stance detection

The paper presents new annotated corpora for performing stance detection on Spanish Twitter data, most notably Health-related tweets. The objectives of this research are threefold: (1) to develop a manually annotated benchmark corpus for emotion recognition taking into account different variants of Spanish in social posts; (2) to evaluate the efficiency of semi-supervised models for extending such corpus with unlabelled posts; and (3) to describe such short text corpora via specialised topic modelling.A corpus of 2,801 tweets about COVID-19 vaccination was annotated by three native speakers to be in favour (904), against (674) or neither (1,223) with a 0.725 Fleiss’ kappa score. Results show that the self-training method with SVM base estimator can alleviate annotation work while ensuring high model performance. The self-training model outperformed the other approaches and produced a corpus of 11,204 tweets with a macro averaged f1 score of 0.94. The combination of sentence-level deep learning embeddings and density-based clustering was applied to explore the contents of both corpora. Topic quality was measured in terms of the trustworthiness and the validation index.     

A Recursive tree-structured neural network with goal forgetting and information aggregation for solving math word problems

Most existing state-of-the-art neural network models for math word problems use the Goal-driven Tree-Structured decoder (GTS) to generate expression trees. However, we found that GTS does not provide good predictions for longer expressions, mainly because it does not capture the relationships among the goal vectors of each node in the expression tree and ignores the position order of the nodes before and after the operator. In this paper, we propose a novel Recursive tree-structured neural network with Goal Forgetting and information aggregation (RGFNet) to address these limits. The goal forgetting and information aggregation module is based on ordinary differential equations (ODEs) and we use it to build a sub-goal information feedback neural network (SGIFNet). Unlike GTS, which uses two-layer gated-feedforward networks to generate goal vectors, we introduce a novel sub-goal generation module. The sub-goal generation module could capture the relationship among the related nodes (e.g. parent nodes, sibling nodes) using attention mechanism. Experimental results on two large public datasets i.e. Math23K and Ape-clean show that our tree-structured model outperforms the state-of-the-art models and obtains answer accuracy over 86%. Furthermore, the performance on long-expression problems is promising.¹     

Class-aware tensor factorization for multi-relational classification

In this paper, we propose a tensor factorization method, called CLASS-RESCAL, which associates the class labels of data samples with their latent representations. Specifically, we extend RESCAL to produce a semi-supervised factorization method that combines a classification error term with the standard factor optimization process. CLASS-RESCAL assimilates information from all the relations of the tensor, while also taking into account classification performance. This procedure forces the data samples within the same class to have similar latent representations. Experimental results on several real-world social network data indicate this is a promising approach for multi-relational classification tasks.     

Information resources estimation for accurate distribution-based concept drift detection

Machine learning applications must continually utilize label information from the data stream to detect concept drift and adapt to the dynamic behavior. Due to the computational expensiveness of label information, it is impractical to assume that the data stream is fully labeled. Therefore, much research focusing on semi-supervised concept drift detection has been proposed. Despite the large research effort in the literature, there is a lack of analysis on the information resources required with the achievable concept drift detection accuracy. Hence, this paper aims to answer the unexplored research question of “How many labeled samples are required to detect concept drift accurately?” by proposing an analytical framework to analyze and estimate the information resources required to detect concept drift accurately. Specifically, this paper disintegrates the distribution-based concept drift detection task into a learning task and a dissimilarity measurement task for independent analyses. The analyses results are then correlated to estimate the required number of labels within a set of data samples to detect concept drift accurately. The proximity of the information resources estimation is evaluated empirically, where the results suggest that the estimation is accurate with high amount of information resources provided. Additionally, estimation results of a state-of-the-art method and a benchmark data set are reported to show the applicability of the estimation by proposed analytical framework within benchmarked environments. In general, the estimation from the proposed analytical framework can serve as guidance in designing systems with limited information resources. This paper also hopes to assist in identifying research gaps and inspiring new research ideas regarding the analysis of the amount of information resources required for accurate concept drift detection.     

Semi-supervised document retrieval

This paper proposes a new machine learning method for constructing ranking models in document retrieval. The method, which is referred to as SSRank, aims to use the advantages of both the traditional Information Retrieval (IR) methods and the supervised learning methods for IR proposed recently. The advantages include the use of limited amount of labeled data and rich model representation. To do so, the method adopts a semi-supervised learning framework in ranking model construction. Specifically, given a small number of labeled documents with respect to some queries, the method effectively labels the unlabeled documents for the queries. It then uses all the labeled data to train a machine learning model (in our case, Neural Network). In the data labeling, the method also makes use of a traditional IR model (in our case, BM25). A stopping criterion based on machine learning theory is given for the data labeling process. Experimental results on three benchmark datasets and one web search dataset indicate that SSRank consistently and almost always significantly outperforms the baseline methods (unsupervised and supervised learning methods), given the same amount of labeled data. This is because SSRank can effectively leverage the use of unlabeled data in learning.     

钢板表面缺陷深度主动学习高效分类方法

针对钢板表面缺陷图像分类传统深度学习算法中需要大量标签数据的问题,提出一种基于主动学习的高效分类方法。该方法包含一个轻量级的卷积神经网络和一个基于不确定性的主动学习样本筛选策略。神经网络采用简化的convolutional base进行特征提取,然后用全局池化层替换掉传统密集连接分类器中的隐藏层来减轻过拟合。为了更好的衡量模型对未标签图像样本所属类别的不确定性,首先将未标签图像样本传入到用标签图像样本训练好的模型,得到模型对每一个未标签样本关于标签的概率分布(probability distribution over classes, PDC),然后用此模型对标签样本进行预测并得到模型对每个标签的平均PDC。将两类分布的KL-divergence值作为不确定性指标来筛选未标签图像进行人工标注。根据在NEU-CLS开源缺陷数据集上的对比实验,该方法可以通过44%的标签数据实现97%的准确率,极大降低标注成本。     

Irony detection via sentiment-based transfer learning

Irony as a literary technique is widely used in online texts such as Twitter posts. Accurate irony detection is crucial for tasks such as effective sentiment analysis. A text’s ironic intent is defined by its context incongruity. For example in the phrase “I love being ignored”, the irony is defined by the incongruity between the positive word “love” and the negative context of “being ignored”. Existing studies mostly formulate irony detection as a standard supervised learning text categorization task, relying on explicit expressions for detecting context incongruity. In this paper we formulate irony detection instead as a transfer learning task where supervised learning on irony labeled text is enriched with knowledge transferred from external sentiment analysis resources. Importantly, we focus on identifying the hidden, implicit incongruity without relying on explicit incongruity expressions, as in “I like to think of myself as a broken down Justin Bieber – my philosophy professor.” We propose three transfer learning-based approaches to using sentiment knowledge to improve the attention mechanism of recurrent neural models for capturing hidden patterns for incongruity. Our main findings are: (1) Using sentiment knowledge from external resources is a very effective approach to improving irony detection; (2) For detecting implicit incongruity, transferring deep sentiment features seems to be the most effective way. Experiments show that our proposed models outperform state-of-the-art neural models for irony detection.