Abstract

Rapid advancements in networks and computer systems have opened a newdoor for immoral acts like cybercrime, which threaten public safety, and security, aswell as the global economy. The purpose of this proposal is to analyse IP fraud andcyberbullying as two distinct types of cybercrime. The primary goals of this study areto use instances of cybercrime to provide a short examination of cybercrime activities,and the family member principles, and propose a pairing schema. Using the NaiveBayes (NB) & Support Vector Machine (SVM) artificial intelligence techniques,cybercrime instances are categorised according to their ideal qualities. The Twitter datain the Kaggle database has been clustered using K-means. User ID, sign-up date,referral, browser, gender, and age as well as IP address are just a few of the most usefulinformation used to educate the computer. Total 151,113 datasets were used forexperimental analysis of the suggested algorithm's performance. The accuracy of thesuggested approach, 97%, is higher than that of the current method (NB). Thechallenge of regression may be easily surmounted with the use of the random forestmethod for the categorization of the resultant cybercrimes. The planned study uses agecategories as the foundation for identifying the different offenses.

Abstract

This article examines the features of the dealer's brush list evaluationmaterial in light of research findings on misleading assessment and identification ofonline purchasing. A Gated Recurrent Unit (GRU) model using keyword weighting ispresented as a solution to the issue that it is challenging for the DL model to collect thefeature data of the whole assessment text in a false evaluation identifying job. The TFIDF technique is first used to generate the list of keywords, and then that list's weight isapplied to the word vector. Finally, a weighted vector of words is categorised using thismethod of the model to finish the recognition job of erroneous evaluation, replacing thepooling component of the GRU model with a constrained Boltzmann machine. Byusing a variety of text categorization algorithms and comparing their results in terms ofcorrectness and performance, this research aspires to represent the practical benefits ofapplications that use machine learning in the real world. We built a system that can runseveral text classification algorithms, and we used that system to create models thatwere educated using actual data taken from E-Commerce, a virtual fashion e-commerceplatform. The Convolutional Neural Network technique achieved the greatest meanaccuracy of 96.08% (with a range of 85.44% to 99.99%) with an average deviation of5.65%.

Abstract

The purpose of text categorization, a machine learning technique, is toautomatically assign tags or categories to texts. Natural language processing (NLP)-based text classifiers can quickly analyse vast volumes of text and classify it based onemotions, themes, and human intent. FastText was created by Facebook's AI Researchteam and is available to the public as a free library. Its primary goal is the efficient andaccurate processing of big datasets in order to provide scalable remedies for theproblems of text categorization and representation. Traditional machine learningtechniques used in most text categorization models suffer from issues including thecurse of dimensionality and subpar performance. This research offers a fastText-basedAI text classification model to address the aforementioned issues. The fastTextapproach allows our model to create a low-dimensional, continuous, and high-qualityrepresentation of text by mining the text for relevant information through featureengineering. The experiment uses Python to define the text dataset, and the resultsdemonstrate that our model outperforms the baseline model trained using classic MLmethods in terms of accuracy, recall, and F values.

Abstract

The rate at which technology is improving is increasing all throughout theworld. Every day, a tremendous amount of textual data is produced as a result of theInternet, websites, business data, medical information, and the media. Extraction ofinteresting patterns from text data with varied lengths such as views, summaries, andfacts is a challenging issue. This work provides a deep learning (DL) algorithm-basedapproach to news text classification to address the issues of large amounts of text dataand cumbersome features obtaining value in news. Although the relationship amongwords as well as categories has a significant impact on the categorization of news text,previous approaches to text classification relied solely on the knowledge of theconnections between words to make their classification decisions. This research usesthe idea of a tailored algorithm to provide a CNN, LSTM, and MLP-basedcustomizable ensemble framework for categorising news text data. The proposedmodel is based on a parallel representation of word vectors and word dispersion. Wefeed the term vector to the CNN module to convey the relationship between words, aswell as nourish the discrete vector corresponding to the relationship between words andcategories into the MLP module to achieve deep learning of the spatial data on features,the time-series feature information, and the connection words and categories in newstexts. Extensive experimental study confirmed the dependability and efficacy of theproposed approach. The experimental results demonstrated that the proposed methodimproved the most - in terms of precision, recall rate, and comprehensive value, whilealso addressing the problems of text length, extraction of features issues with the newstext, and classification of news text.

Abstract

The unique Coronavirus pandemic of 2019 (called COVID-19 by the globeHealth Organisation) has exposed several governments throughout the globe to risk.The Covid-19 epidemic, which had previously only affected the Chinese population, isnow a major worry for countries all over the globe. Additionally to the obvious healtheffects of COVID-19 epidemic, this study reveals its repercussions on the worldwideeconomy. The research went on to talk about how they analysed public opinion andlearned new things about Covid-19 vaccinations by using content Analytics andsentiment evaluation in Natural Language Processing (NLP) using content fromTwitter. To categorise and analyse the outcomes, researchers used two machinelearning algorithms: logistic regression (LR), random forest, decision tree, andconvolutional neural networks (CNNs). To better identify public opinion, several preprocessing methods were used and categorised responses into neutral, positive, andnegative categories. The public's opinion on Covid-19 vaccinations is 31% favourable,22% negative, and 47% neutral, according to the results of the emotion sectiondistribution. CNN achieved 98% accuracy, according to the tested machine learningalgorithms.

Abstract

The use of sarcasm in everyday conversation has recently risen toprominence. All the kids of our age utilise sarcasm to convey a negative message in amore nuanced manner. With the advancement of AI and machine learning techniquesin the area of natural language processing (NLP), it has become more difficult toreliably and effectively identify sarcasm. This research provides a new method forsarcasm detection using machine studying and deep learning, hoping to make ameaningful contribution to this expanding area of study. In order to prepare the phrasefor a hybrid deep learning model for training and classification, this method employsbidirectional encoder representations from transformers (BERT). The combination ofCNN and LSTM creates the hybrid model used here. The suggested model has beentested on two datasets, with the goal of identifying sarcasm from nonsarcastic words.The trained model obtained a 99.63% accuracy rate, a 99.33% precision rate, a 99.83%recall rate, and an F1-score of 99.56%. These findings are based on 10 rounds of crossvalidation performed on the model that was suggested using the medical datasets. 

Abstract

Classifying texts into groups determined by their content is called textclassification. In this process, automatic labelling of documents written in naturallanguages is carried out according to predetermined labels. Both text comprehensionsystems, which perform transformations on texts such as creating summaries,answering queries, and extracting data, and text-retrieving systems, which obtain textsin fulfillment of a user query, rely heavily on text categorization. In order to learneffectively, current algorithms for supervised learning for text classification need alarge enough training set. This research introduces a novel text categorizationalgorithm that uses artificial intelligence techniques (machine studying and deeplearning techniques) and needs fewer documents for training than previous methods.To generate a feature set from already-classified text documents, we resort to “wordrelation,” or association rules based on these words. To classify the data, we use theidea of a Convolutional Neural Network with Deep Convolution to the extractedfeatures and then employ a single genetic algorithm approach. The suggested methodhas been built and thoroughly tested in a working system. The results of theexperiments show that the suggested system is an effective text classifier.

Abstract

When it comes to training ML systems, internet-based data is invaluable.Despite the difficulty in collecting this information, teams of experts from academicinstitutions and research labs have created publicly accessible databases. Twitter andother social media platforms provided large quantities of useful information throughoutthe pandemic, which was used to evaluate healthcare decisions. In order to forecastillness prevalence and offer early warnings, we suggest analysing user attitudes byusing efficient supervised machine learning algorithms. The gathered tweets weresorted into positive, negative, and neutral categories for preprocessing. Hybrid featureextraction is the innovative aspect of our work; we used it to correctly describe posts bycombining syntactic features (TF-IDF) and semantic elements (FastText and Glove),which in turn improved classification. The experimental findings suggest that whenusing Naive Bayes, the combination of FastText and TF-IDF achieves the best results.

Abstract

Effectively classifying texts is possible using several classificationtechniques. Machine learning constructs a classifier by studying and memorising thecharacteristics of several classes. For text categorization, deep learning provides similaradvantages since it can function with great precision using simpler architecture andprocessing. In order to categorise textual information, this research makes use ofmachine learning and deep learning methods. There is a great deal of extraneous detailsin textual data that must be removed during pre-processing. To prepare it for analysis,we remove duplicate columns and impute missing data. In the next step, we use deeplearning techniques for classification, including long short-term memory (LSTM),artificial neural network (ANN), as well as gated recurrent unit (GRU). According tothe findings, GRU obtains 92% accuracy, which is higher than that of any other modelor baseline investigation.

Abstract

 In text categorization, texts are sorted into groups according to their content.It is the process of automatically classifying texts written in natural languagesaccording to a set of guidelines. Both text comprehension systems, which performtransformations on text such as creating summaries, answering queries, and extractingdata, and retrieval of text systems, which collect texts in reaction to a user query on theinternet content, rely heavily on text categorization. In order to learn effectively,current algorithms for supervised learning for text classification need a large enoughtraining set. This research introduces a novel text categorization system that makes useof an AI approach and needs fewer articles for training over information found on theweb. To generate a feature set from already categorised texts, we resort to “wordrelation,” or association rules based on these terms. The obtained characteristics arethen processed by a Support Vector Machine, and ultimately, a single genetic algorithmidea is introduced for classification. The suggested approach has been developed andvalidated in a working system. The experimental results verify the effectiveness of theproposed system as a text classifier.

Abstract

The study of emotions has emerged as an important area of research becauseof the wealth of information it can provide. Emotions can be expressed in a variety ofways including words, facial expressions, written material, and movements. Naturallanguage processing (NLP) & deep learning concepts are essential to solving thecontent-based classification problem that is emotion detection in a text document.Therefore, in this research, we suggest using deep learning to aid semantic text analysisin the task of identifying human emotions from transcripts of spoken language. Visualforms of expression, such as makeover jargon, may be used to convey the feeling.Datasets of recorded voices from people with Autism Spectrum Disorder (ASD) aretranscribed for analysis. However, in this paper, we specialize in detecting emotionsfrom all of the textual dataset and using the semantic data enhancement process to fill afew of the phrases, or half-broken speech, as patients with Autism Spectrum Disorder(ASD) lack social contact skills due to the patient not very well articulating theircommunication.

Abstract

Herein, we introduce a method for website classification using deep neuralnetworks and mixed data extractors. We use iterative training as well as supervisedlearning approaches to use a gradient descent methodology to simulate the websitecategorization. This modern model is comprised of a webpage encoder, a convolutionalneural network (CNN) feature extraction, a bidirectional long short-term memory(LSTM) feature extractor, as well as a fully connected classifier. It may retrieve variouswebsite features at various granularities. Our model may quickly select a suitablewebsite class by concatenating mixed features obtained from mixed feature extractors.On the realistic website dataset that has been obtained, we conduct in-depth tests. Thedataset is compiled using domains that were taken from the telecom operator's DNSrecords. The proposed categorization schema outperforms state-of-the-art models incomparison to our fresh model as well as a slew of popular machine learningalgorithms in terms of accuracy, recall, F1, and precision. Other web apps may benefitfrom all of this as well, such as detecting fake websites as well as ads.

Abstract

Text classification is a difficult technique. Many techniques have beendeveloped to decrease the dimension of feature vectors for use in text classification dueto their enormous size. This work provides a detailed discussion of unique parametersutilising an optic clustering strategy, as well as a review of some of the most essentialtext categorization algorithms. In this case, the words are clustered according to theirlevel of similarity. Each cluster's membership function is based on the mean along withthe standard deviation of its data. Finally, characteristics are chosen from eachgrouping. Each cluster's extracted feature is the weighted sum of its words. There's alsono need to guess or use trial-and-error approaches to determine the optimal number ofclusters.

Abstract

An integral aspect of natural language processing is text categorization, thegoal of which is to assign a predetermined category to a given text. Feature selectionand categorization models come in a wide variety of forms. Most researchers, however,would rather utilise the prepackaged functions of existing libraries. In the field ofnatural language processing (NLP), automated medical text categorization is veryhelpful for decoding the information hidden in clinical descriptions. Machine learningapproaches seem to be fairly successful for medical text categorization problems;nevertheless, substantial human work is required in order to provide labelled trainingdata. Clinical and translational research has benefited greatly from the computerisedcollection of vast amounts of precise patient information, including illness status, bloodtests, medications taken, and side effects, along with therapy results. As a result, themedical literature contains a massive amount of information on individual patients,making it very difficult to digest. In this research, we suggest using N-grams and aSupport Vector Machine (SVM) to classify healthcare-related texts. We conductexperiments to determine the viability of our code and analyse it across a variety ofcategorization methods.

Abstract

When applying conventional text classification techniques, the values inagricultural text are converted into characters, which destroys the original semanticrepresentation of numerical aspects. A unique text classification approach is suggested,based on the dynamic fusion of several characteristics, to completely mine the deeplatent semantic characteristics in agricultural literature. The global key semanticcharacteristics of the text were extracted using the Bi-directional Gated RecurrentNeural Networks (GRU) model with attention mechanism, while the local semanticdata about the text at various levels was extracted using the multiple windowsConvolution Neural Network. Finally, the number that features essential semanticexpressions was obtained using a machine learning approach for creating thequantitative value feature vector. To further enhance the deep semantic expressionfound in agricultural text and successfully improve the impact of farm textcategorization with phenotypic numerical type, we use a focus technique todynamically fuse the derived numerous semantic characteristics.

Abstract

We presented an Information Retrieval (IR) system that learns from existinginformation and creates a single pertinent document that, we presume, has all itsindexed pertinent details for a query. Deep learning makes such a system viable. Wethen asked people to score the query plus word-cloud representation of three randomlyselected relevant texts and our new synthetic document. The synthetic documenttopped all inquiries and users. We then trained a CNN using query-relevant data. Weperformed ”deep learn” function on a synthetic, relevant material using the CNN. Weused crowdsourcing to compare the “deep-learned” material to related documents.Users can see a query and four-word cloud (three relevant documents and our deeplearning synthetic document). The synthetic document provides the the most relevantfeedback. 

Abstract

Lexical model BERT already trained on BookCorpus and Wikipedia workswell on two NLP tasks after downstream fine-tuning. The requirements of BERTmodel include strategy analyses and task-specific and domain-related data. Theproblems of task awareness as well as instruction data in BERT-DL, a BERT-basedtext-classification model, are addressed through auxiliary sentences. The pre-training,training, and post-training steps for BERT4TC's domain challenges are all provided.Learning speed, sequencing duration, and secret state vectors that select fine-tuning areall investigated in extended trials over 7 public datasets. The BERT4TC model is thencontrasted using a variety of auxiliary terms and post-training goals. On multiple-classdatasets, BERT4TC with the ideal auxiliary phrase outperforms previous state-of-theart feature-based algorithms and fine-tuning approaches. Our domain-related corpustrained BERT4TC beats BERT on binary sentiment categorization datasets.

Abstract

Supervised learning based on deep learning is often used for mass-scalepicture categorization. However, it takes a lot of computing effort and energy to retrainthese vast networks to accept new, unknown data. When retraining, it is possible thattraining samples used before would not be accessible. We present a scalable, graduallyexpanding CNN that can learn new jobs while reusing some of the base networks andan efficient training mechanism. Our approach takes cues from transfer learningmethods, but unlike other approaches, it retains knowledge of previously masteredtasks. Convolutional layers from the early section of the base network are reused in theupdated network, and a few more convolutional kernels are added to the later layers tofacilitate learning a new set of classes. On the task of categorising texts, we tested thesuggested method. Our method achieves comparable classification accuracy to thestandard incremental learning method in which networks are updated solely with newtraining samples, without any network sharing), while also being more resourcefriendly and taking less time and space to train.

Abstract

More individuals are sharing their thoughts and feelings through internetvideos as social media platforms proliferate. While successful emotional fusion inmultimodal data is a key component of multimodal sentiment analysis, most existingresearch falls short in this area. Predicting users' emotional inclinations through theirexpressions of language is made easier by multi-modal sentiment detection. As a result,the field of multi-modal sentiment detection has grown rapidly in recent years. As aresult, multimodal sentiment analysis is quickly rising to the forefront of academicinterest. However, in actual social media, visuals and sentences do not always worktogether to represent emotional polarity. Additionally, there are several informationmodalities, each contributing in its unique way to the overall emotional polarity. Amultimodal approach to sentiment analysis that takes into account contextualknowledge is presented as a solution to these issues. The approach begins by miningsocial media texts for subject information that comprehensively describes the commentmaterial. Additionally, we use cutting-edge pre-training models to identify emotionalqualities that span several domains. Then, we provide methods for merging features atdifferent levels, such as cross-modal global fusion as well as cross-modal high-levelsemantics fusing. At long last, we run our tests on a multimodal dataset that reallyexists in the real world. Results show that the proposed approach can correctly classifythe tone of heterogeneous online reviews, and it also outperforms the standardapproach in many other ways as well.

Abstract

A natural catastrophe is an example of a rare occurrence that does nothappen often but may have devastating effects on people and their environment when itoccurs. People now have a quick and easy outlet for voicing their ideas thanks to socialmedia. Thus, it may be used by researchers to learn about how individuals react to andthink about a wide variety of extremely unusual occurrences. Many research works usesocial media data to investigate how people's reactions to unusual occurrences in thereal world translate to their online personas, thoughts, feelings, and actions. In thispiece of work, we offer a method for extracting features and classifying tweets onunusual events like Hurricane Sandy. To begin, a new approach to feature extraction ispresented, one that may be used to extract relevant features from each communication.The next step is to offer a Score-based categorization system for differentiatingbetween communications about events and those that are unrelated. Finally, thedevelopment of a rare event is analyzed using our suggested approach and the popularkeyword search method. The findings show that the suggested method is effective indistinguishing between text messages connected to unusual events and those that areunrelated.

Abstract

Text categorization is a crucial technology in data mining as well as dataretrieval that has been extensively investigated and is developing at a rapid pace.Convolutional neural networks (CNNs) are a kind of deep learning modeling that mayreduce the complexity of the model while accurately extracting characteristics frominput text. Support vector machine (SVM) results have always been more trustworthyand superior to those of other traditional artificial intelligence approaches. Usingenhanced convolutional neural network (CNNs) as well as support vector machines(SVMs), we offer a novel approach to online text categorization in this study. Ourapproach begins with text attribute identification and prediction using a model based onCNN with a five-layer network structure. Databases including both text and imageswill find it to be a major factor in the long run. 

Abstract

Increasing numbers of people are using the Internet to share their feelingsand communicate with one another, and the vast majority of these expressions ofemotion take the form of text. Using sentiment dictionaries, machine learning, and deeplearning are the three most common approaches to text sentiment categorizationstudies. Due to the exponential growth of textual data, it is crucial to create models thatcan automatically analyse this material. Labels like gender, age, nationality, emotion,etc., may be included in the texts. Numerous investigations of text categorization haveemerged because the use of such labels may be useful in various commercial sectors.The Convolutional Neural Network, also referred to as CNN, was recently utilised tothe problem of text categorization, with promising results. In this study, we advocatefor the use of convolutional neural network networks for the job of classifyingemotions. Using three popular datasets, we demonstrate that our networks outperformexisting cutting-edge deep learning models by using successive convolutional layers toprocess substantially longer sentences.

Abstract

One crucial task in text processing is determining how similar two papersare to one another. A novel similarity metric is suggested in this study. Finding asuitable similarity metric for written materials that permits the development of coherentgroupings is a significant difficulty for document clustering. After that, we use TFIDFto build a vector space, and then we use the ward's approach and the K-meansalgorithm to accomplish clustering. WordNet is additionally employed in the process ofsemantic document clustering. Visualisations and an interactive website illustrating theconnections between all clusters illustrate the findings. The existence (and quantity) ofwords in texts are all that are taken into account while utilising the traditional bag-ofwords paradigm. This process might lead to texts with identical meanings but distinctvocabulary being placed in various groups. The findings acquired using the suggestedapproach are analysed for their correctness using the F-measure. Comparisons usingthe sentence vectors model (Doc2vec) and the bag-of-words model are made toconfirm the edge of the suggested strategy. The suggested methodology may be used todecipher web chat logs and client feedback posted online. We evaluate our method on avariety of real-world data sets including examples of text classification and clusteringproblems. The findings prove that the proposed measure outperforms competingstrategies.

Abstract

During times of disaster or epidemic, social media has emerged as a vitalmeans of communication. It is difficult to examine the complete situational awarenessvia many elements and emotions to aid authorities due to the unpredictability of thesecalamities. Currently, systems for aspect recognition and sentiment analysis relyheavily on labelled data and require human curation of aspect categories. To analyzepublic opinion from a variety of angles, this study suggested a hybrid text analyticalapproach. Using the popular Latent Dirichlet Allocation (LDA) topic modeling, wefirst extracted and clustered the elements from the data. We then used the linguisticinquiry and word count (LIWC) lexicon to extract the sentiments and label the dataset.Finally, in the third layer of our structure, we mapped the elements into emotions, andthe sentiments were classified using well-known machine learning classifiers. Thecomparison of our technique with other aspect-oriented sentiment analysis approachesshows encouraging results in experiments with actual datasets, and our method withseveral variants of classifiers surpasses current methods with top F1 scores of 91%.

Abstract

When dealing with data in the form of text, the most popular method fordoing analysis and determining sentiment content is called “Sentiment Analysis.”Sentiment analysis is also known as Opinion Mining. Suggestions, feedback, tweets,and comments are all examples of the various types of text data that are being created.Customer feedback on e-commerce sites is a constant source of new information.Online stores may better meet client needs, improve their services, and boost sales byanalyzing E-Commerce data. Positive, negative, and neutral feedback from customersmay be separated using sentiment analysis. Numerous methods for Sentiment Analysishave been developed by academics. Typically, only a single machine learningalgorithm is used for sentiment analysis. The purpose of this study, which makes use ofAmazon review data, is to extract positive, negative, and neutral review ratings bylocating aspect phrases, identifying the Parts-of-Speech, and applying classificationalgorithms to the collected data.

Abstract

There are now a great deal of consumer reviews of items that are writtenentirely in text. The reviews express their opinions. Opinion mining is another name forsentiment analysis. A common way for businesses to keep tabs on how customers feelabout their brands and goods is through the use of sentiment analysis on textual data.Naive Bayes, Random Forest, Decision Tree, and Support Vector Machine classifiersare all used and compared in this study. In this study, we evaluate the efficacy ofseveral classifiers by measuring their ability to correctly categorize mobile product datasets of varying sizes. Data were collected from popular online retailers like Amazon,Flipkart, and Snapdeal and analyzed to determine categorization accuracy. NaiveBayes, Random Forest, Decision Tree, and Support Vector Machines are some of thecategorization algorithms compared here.

Abstract

The identification of problems has become easier in sentiment categorizationusing conventional neural network–based short text classification methods . Word2vec,a convolutional neural network (CNN), and Bidirectional Long-term and Short-termMemory networks (LSTM) are used incombination to overcome this issue. UsingWord2vec word embeddings, the CNN-LSTM model was able to attain an accuracy of91.48%, as demonstrated experimentally. This demonstrates that the hybrid networkmodel outperforms the single-structure neural network when dealing with relativelybrief texts.

Abstract

Sentiment analysis, the practice of classifying and identifying opinionsdisplayed in audio, words, database reports, and tweets to ascertain if the opinion ispositive, neutral, or negative, is of great interest to many individuals in themicroblogging service arena. It could be challenging to extract sentiment from Twitterdata due to its quirks. The research suggests a way to analyze sentiment using aHadoop infrastructure with a deep learning classifier. In order to gather characteristics,data is distributed across Hadoop nodes. After that, the data from Twitter is parsed forthe most crucial parts. The input data from Twitter is sorted into two categories,positive review along with negative review, via a deep learning algorithm, including adeep recurrent neural networks classifier. Some of the metrics used to measureperformance include classification accuracy, specificity, and sensitivity. With asensitivity of 0.9404 and a generality of 0.9157, the proposed technique surpassedtraditional methods in classification with a precision of 0.9302. 

Abstract

A novel approach to creating a dataset to train a neural network that cananalyze the tone of social media postings is proposed in this study. The paper goes onto detail how the word2vec and BERT algorithms may be used in a neural network toanalyze social media messages and evaluate their emotional tone. A hybrid of cosinesimilarity as well as ontological mappings based on a tweaked version of the TermFrequency-Inverse Document Frequency (TFIDF) characteristics is used by thealgorithm for semantic searching. The execution includes sentiment analysis, phraseextraction process, textual belief indicator, keyword-based search, as well as textsummary, among other things. Additionally, trials were carried out proving the efficacyof the methods presented. The efficiency of stemming and lemmatization of the text increating a training set for sentiment analysis was also tested experimentally. 

Abstract

Text-based web archives have become increasingly common as technologyhas advanced. For many text classification applications, classic machine learningclassifiers like support vector machines (SVMs) and naïve Bayes (NBayes) performwell. Since short texts have fewer words and convolutional and pooling layers havetheir limits, these classifiers suffer from sparsity and lack long-term dependencies. Inthis study, we present a convolutional recurrent neural network architecture that makesuse of a modified skip-gram method. For the adversarial training of the skip-gramalgorithm, we employ the L2 regularization technique. It can boost the model'sperformance in text sentiment classification tasks and increase its robustness andgeneralizability. To extract information from the entire text while dampening theinfluence of irrelevant words, we deployed a convolutional neural network equippedwith attention mechanisms. The CNN-based categorization of text emotion is complete.When compared to other classifiers used on the Twitter dataset, our model andalgorithm were shown to be more efficient and accurate.

Abstract

More and more people are inclined to use GIFs, videos, and photographs onsocial media as a way to convey their feelings and thoughts. We developed a Pythonbased multimodal sentiment analysis tool for various Twitter formats, taking intoaccount not just the text of a tweet but also its accompanying GIFs and pictures, formore precise sentiment scoring. We employ fine-tuned CNN for image sentimentanalysis, VADER for text analysis, and image sentiment and facial expression analysisfor GIFs, with each frame individually analyzed. Our research shows that combiningtextual and picture data yields superior outcomes compared to models that depend onlyon either images or text. The output scores from our text, picture, and GIF modules willbe aggregated to get the final sentiment score for the incoming tweets.

Abstract

As a result of the COVID-19 epidemic, many individuals are experiencingextreme worry, dread, and other difficult emotions. Since coronavirus immunizationswere first introduced, people's reactions have gotten more nuanced and varied. In thisstudy, we will use deep learning methods to decode their emotions. Twitter provides aglimpse into what is popular and what is on people's minds at any given time, andsocial media is presently the finest means to convey sentiments and emotions. Our goalwhile conducting this study was to have a better grasp of how different groups ofindividuals feel about vaccinations. The research period for this research's tweet wasfrom December 21st to July 21st. Of the most talked-about vaccines that have recentlybeen available in various regions of the world were the subject of several tweets. Theterm Valence Aware Sentiment Dictionary An NLP program called Believed(VADER) was used to examine people's sentiments on certain vaccines. We werebetter able to see the big picture after categorizing the collected attitudes into positive(33.96 percent), negative (17.55 percent), and neutral (48.49 percent) camps. We alsoincluded into our study an examination of the tweets' chronology, given that attitudeschanged over time. The performance of the forecasting models was evaluated using anRNN-oriented design that included bidirectional LSTM (Bi-LSTM) as well as longshort-term memory (LSTM); LSTM attained an accuracy of 90.59% as well as BiLSTM of 90.83%. Additional performance metrics, such as Precision, F1-score, as wellas a matrix of confusion, were used to confirm our hypotheses as well as outcomes.The findings of this research provide credence to efforts to eradicate coronavirus acrossthe globe by expanding our knowledge of public opinion on COVID-19 vaccines. 

Abstract

Twitter, Facebook, Instagram, etc. are just a few of the many onlinediscussion platforms that have sprung up as a result of the explosion in internet use andpopularity, giving individuals a place to air their views on current events. Films getboth acclamation and criticism from the general public. As a major form ofentertainment, they inspire user evaluations of film and television on websites likeIMDB and Amazon. Scientists and researchers give careful thought to these critiquesand comments in order to extract useful information from the data. This data lacksorganisation but is of critical importance nevertheless. Opinion mining, also known assentiment classification, is a growing field that uses machine learning and deeplearning to analyse the polarity of the feelings expressed in a review. Since texttypically carries rich semantics useful for analysis, sentiment analysis has grown intothe most active investigation in NLP (natural language processing). The continuousprogress of deep learning has substantially increased the capacity to analyse thiscontent. Convolutional Neural Networks (CNN) are commonly utilised for naturallanguage processing since they are one of the most successful deep learningmethodologies. This paper elaborates on the methods, datasets, outcomes, and limits ofCNN-based sentiment analysis of film critics' reviews.

Abstract

In our research, we use sentiment analysis to determine how well ratings andreviews are compared on Amazon.com. The process of determining whether a text'stone is favorable or negative and labelling it as such is known as sentiment analysis.Consumers may write evaluations on e-commerce sites like Amazon.com and indicatethe polarity of their opinion. There is a discrepancy between the review and the ratingin certain cases. We used deep learning to analyze the sentiment of Amazon.comproduct reviews in order to find reviews with inconsistent star ratings. A paragraphvector was utilized to transform textual product evaluations into numeric data that wasthen fed into a neural network with recurrent equipped with a gated recurrent unit fortraining. We built a model that takes into account the review text's semanticconnections to the product data. Additionally, we built a web service that uses thetrained model to predict the rating score of a submitted review and gives feedback to areviewer if the anticipated and submitted ratings do not line up.

Abstract

 Many hotel evaluations have been written and shared online these days.Machine learning sentiment classification requires complicated artificial designfeatures and feature extraction technique, whereas emotion dictionary-based sentimentclassification requires a large amount of emotional database resources. In this study, wepresent the idea of long short-term memory. The text categorization method is used todetermine the general tone. First, the brief comment text is processed into the LSTMnetwork using word2vec and word segmentation technology; next, a dropout techniqueis implemented to avoid overfitting in order to get the final rating model. By using theLSTM network's superior short-term memory, a positive impact has been realized onsentiment categorization of reviews of hotels, with a precision of more than 95%.

Abstract

Retailers, market analysts, and other users of the web are greatly influencedby user views. Arranging the unstructured data gathered from various social medianetworks correctly is necessary for doing relevant analysis. Emotional evaluation as amethod for cross-lingual data classification has received considerable attention. Textualorganization is a subfield of natural language processing, or NLP, that may be used toclassify an individual's emotional or mental condition as positive, negative, beneficial,or detrimental, like a thumbs up or thumbs down, etc. A combination of sentimentanalysis as well as deep learning techniques might be the key to solving this kind ofproblem. Deep learning models, which are capable of machine learning, areparticularly useful for this. One of the most widely used deep learning architectures foranalyzing sentiment in text is called Long Short Term Memory. These frameworkshave potential applications in NLP. In this study, we provide algorithms to solve theproblem of multilingual sentiment analysis, and we evaluate their precision factors todetermine which one is the most effective.

Abstract

The amount of textual information on the internet has increased significantlywith the debut of social media platforms like Twitter, including news stories andhistorical records. This is due to the growth of Web 2.0. More individuals are using theinternet and different forms of social media to share their thoughts and feelings withthe world. As a result, more phrases with emotional nuance were created by the generalpublic. It is natural that researchers will look into new approaches to understandingpeople's emotions and reactions. In addition to providing a novel hybrid system thatcombines text mining and neural networks for sentiment categorization, this studyevaluates the efficacy of many machine learning and deep learning techniques. Morethan a million tweets from across five different topics were utilized to create thisdataset. Seventy-five percent of the dataset was used for training, while the remainingtwenty-five percent was used for testing. When compared to traditional supervisedlearning methods, the system's hybrid approach displays a maximum accuracy of83.7%.

Abstract

The proliferation of Web 2.0 has resulted in a deluge of real-time,unstructured data such as user comments, opinions, and likes. The lack of structure inthe data makes it difficult to create a reliable prediction model for sentiment analysis.There have been promising applications of several DNN architectures to sentimentanalysis, however, these methods tend to treat all features identically and struggle withhigh-dimensional feature spaces. In addition, existing techniques fail to effectivelycombine semantic and sentiment knowledge for the purpose of extracting meaningfulrelevant contextual sentiment characteristics. This paper proposes an integratedconvolutional neural network, or CNN, architecture that takes sentiment as well ascontext into consideration as a means of intelligently developing and highlightingsignificant components of relevant sentiment contextual in the text. To start, we usetransformers' bidirectional encoder representations to create sentiment-enhancedembeddings of words for text semantic extraction using integrated emotion lexiconswith broad coverage for feature identification. The proposed approach then adjusts theCNN in a way that it can detect both word order/contextual text semantics data as wellas the long-dependency relationship in the phrase sequence. Our approach alsoemploys a system to prioritize the most important portions of the phrase sequence. Onelast step in sentiment analysis is the use of support vector machines (SVMs) to reducethe complexity of the space of features and identify locally significant characteristics.The accuracy of existing text sentiment categorization is greatly improved by the use ofthe proposed model, as shown by an evaluation of real-world benchmark datasets.

Abstract

Big Data is here, thanks to the proliferation of social media and IoT devices.As a result of the immense benefit that Big Data has brought to the public as well asbusinesses, the question of how to handle and utilise it more effectively has captivatedpeople from every walk of life. Big Data processing has been plagued by difficultiesdue to the 4V features of Big Data. Solving the data quality issue is crucial to Big Dataprocessing, as is ensuring data quality, which is a precondition for Big Data to play itsworth. Two examples of where Big Data has been put to good use are inrecommendation and prediction systems. In this research, we investigate Big Datarecommendation and prediction systems and endeavour to ascertain the quality of dataat each stage of the process: collection, preprocessing, storage, and analysis. Theproposed remedy follows from a thorough description and examination of the issue athand. We have left a few questions unanswered at the conclusion of the report.

Abstract

Given the real-time and pervasive nature of social media, data mining fortraffic-related insights is a newly emerging area of study. In this work, we discuss thechallenge of mining social media for traffic-generating microblogs on Sina Weibo, theChinese equivalent of Twitter. It is recast as a classification issue in short texts formachine learning. In the first step, we use a dataset of three billion microblogs and thecontinuous bag-of-word model to learn word embedding representations. Wordembedding, as opposed to the standard one-hot vector representation of words, hasbeen shown to be useful in natural language processing tasks because of its ability tocapture semantic similarity between words. Then, we suggest feeding the learned wordembeddings into convolutional neural networks (CNNs), long short-term memory(LSTM) models, and their combination LSTM-CNN to extract traffic-relatedmicroblogs. We evaluate the suggested techniques against state-of-the-art methodssuch as the support vector machine (SVM) model using a bag of n-gram features, theSVM model using word vector features, and the multi-layer perceptron model usingword vector features. The proposed deep learning methods are shown to be useful inexperiments.

Abstract

Users (people/patients) concerned about health concerns have an easy outletin online medical forums along with other public social media on the Internet. TheWorld Health Organization declared a global public health emergency in response tothe emergence of a new coronavirus (infection which causes the disease termedCOVID-19) in late December 2019. In this research, we employed a natural languageprocessing (NLP) technique based on topic modeling to automatically extract COVID19-related talks from social media and discover numerous concerns linked to COVID19 from public viewpoints. As an added bonus, we look into the possibility ofemploying a long short-term memory (LSTM) recurrent neural network to accomplishthe same task with COVID-19 remarks. Our research highlights the value ofincorporating public opinion and appropriate computational approaches into theprocess of learning about and making decisions on COVID-19. The trials also showedthat the study model was able to reach an accuracy of 81.15 percent, which is greaterthan the accuracy attained by many other popular machine-learning methods forCOVID-19 sentiment classification.

Abstract

A student's chances of landing a job after graduation are influenced by theirperformance in school and their history of academic accomplishment. Students whowant to succeed in the working world need to develop skills including technicalproficiency, critical thinking, and effective communication. Here we are making anattempt to figure out how students' academic performance affects their futureopportunities. Algorithms for data mining play a key role in analysing and forecastingthe chance for students' placement according to their previous academic achievement.We polled students at a prominent technical institution for this piece. Multiple factorsthat influence students' probabilities are included in the dataset, and these factors areexamined and shown graphically. We made an effort to assess the data and providevisualisations and insights before running or applying algorithms for machine learningto this dataset. Data analysis, understanding, and preparation are the key focuses of thisresearch. 

Abstract

An essential feature of an intelligent workplace is error-free manufacturing.The data-driven approaches rely on fundamental status data with minimal volume,while the conventional model-based methods rely heavily on accurate equipmentmodels. Not only are these aspects problematic, but they also prevent them frommeeting the real-time need of evaluating industrial big data in an IoT setting. In thisresearch, we provide a fault prediction approach that uses industrial big data to unearththe connection between data (such as status and sound data) and equipment failuresusing machine learning techniques. Not only that but the breakdown could beinvestigated promptly since the equipment's status could be tracked in real-time. Oursolution outperforms the current ones in terms of accuracy and real-time capabilities,according to the simulation results.

Abstract

In order to carry out categorization and generate new categories, the hugevolumes of textual materials produced nowadays must be immediately organised. Thefundamental method of gaining insights from organising textual data is textclassification. Then, we further classify the classes based on the discovered text types.Separated into four stages—pre-treatment, text representation, classifier execution, andclassification—we use a wide range of machine learning approaches to classify texts.In this study, we utilise real-world data from Twitter to evaluate and compare severalsentiment analysis approaches. We clean the data and divide it into train and text setbefore developing models using various vectorising approaches and compare theoutcomes. Based on a comparison of the models with various vectorizations, it wasfound that the best performance was provided by the Logical Regression (LR) modelsusing TF-IDF, with an f1 value of 0.81 and good accuracy and recollection values. 

Abstract

The exponential growth in internet use has also led to the proliferation oftextual information in large quantities. Since handling unstructured material manuallyis difficult, there is a need to explore novel techniques for automated categorization oftextual information. The primary goal of text categorization is to teach a model tocorrectly categorise an unseen text. In this research, the Doc2vec word embeddingtechnique was used to classify stories in Turkish from the TTC-3600 database ofTurkish news and BBC news stories in English. In addition to the CNN based on deeplearning, traditional machine learning classification methods including Gauss NaiveBayes (GNB), Random Forest (RF), Naive Bayes (NB), and Support Vector Machine(SVM) are used. The best classification results using CNN were achieved with theproposed model, scoring 94.17% on the Turkish database and 96.41% on the Englishdatabase.

Abstract

Banks and other financial organisations rely heavily on credit scoring (CS)as a method of risk management since it is both effective and necessary. It reducesfinancial risks and gives sound advice on loan disbursement. As a result, businessesand financial institutions are exploring innovative automated solutions to the CSdilemma in an effort to safeguard their resources and those of their clients. The use ofvarious machine learning (ML) as well as data mining (DM) approaches has led tosignificant progress in CS prediction in recent years. The Deep Genetic HierarchicalNetwork of Learners (DGHNL) is a novel approach developed for this study. SupportVector Machines (SVMs), k-nearest Neighbours (kNNs), Probabilistic NeuralNetworks (PNNs), and fuzzy structures are just some of the many types of methods thatmay be used in the suggested method. The Statlog German (1000 occurrences)approval of credit dataset from the UCI machine learning library was used to evaluateour model. We used a DGHNL model with five unique learner types, two featureextraction methods, three kernel functions, and three methods for optimising modelparameters. In addition to conventional cross-validation (CV) and train-testing(stratified 10-fold) methods, this model employs a cutting-edge biological layeredtraining (participant selection) approach. Using data on German credit approvals fromStatlog, we show that the suggested DGHNL model can obtain a prediction accuracy of94.60% (54 errors per 1000 classifications) with its 29-layer architecture.

Abstract

The practice of “lifelogging” involves documenting an increasing amount ofone's subjective everyday experience with the intention of using the recordings in thefuture as a memory aid or the foundation for data-driven self-development. Therefore,the usefulness of the generated lifelogs depends on the lifeloggers' ability to efficientlysift through them. The logs' intrinsic multi-modality and semi-structure allow them tocombine data from a variety of sources, including cameras and other wearable physicaland virtual sensors. As a result, expressing the data in a graph structure allows for theeffective capturing of all created interrelations. Alternative methods must be developedto capture the higher-level semantics because it is impossible to manually ormechanically annotate each entry with a significant amount of semantic context. Wedescribe an Improved Life Graph (ILG), a first method for building a KnowledgeGraph-based lifelog representation and retrieval solution, which can capture a lifelog ina graph structure and augment it with external data to help with the connection ofhigher-level semantic information.

Abstract

Cardiovascular illness is afflicting enormous monetary and psychologicalcosts. The development of an ASHRO-based model for forecasting healthcare resourceuse and its link with clinical outcomes was driven by a desire to improve the economyand provide a high-quality evaluation of the healthcare system. Data included in thisanalysis were taken from a big database that included doctor visits, insurance claimsacross several years, and results of preventive health screenings. Hospitalized patientswith heart illness (ICD-10 I00-I99) comprised the study population. Broadly definedcomposites compliance served as the explanatory variable, while medical as well aslong-term care costs served as the objective variable. Using a combination of randomforest learning (AI) and multiple regression analysis, predictive models werecalibrated. These models were then used to create ASHRO scores. Two measures, thearea under the curve as well as the Hosmer-Lemeshow test, were used to assess theprediction model's effectiveness. After controlling for clinical risk variables, wecompared the two ASHRO 50% threshold groups' total morbidity at 48 months offollow-up using matching propensity scores. Heart disease affected 61.9% of the48,456 patients surveyed, with an average age of 68.3 9.9 years at hospital release. Forthe purpose of adherence score classification, machine learning was employed tocombine eight factors into a single index: generic drug rate, interconnecting outpatientvisits/clinical laboratory as well as physiological tests, the proportion of daysaddressed, secondary mitigation, rehabilitation magnitude, direction, and a single indexthat adjusted for eight factors. In the end, the multiple regression study yielded a 0.313(p 0.001) coefficient of determination. Medical as well as long-term care expenditureshad a statistically significant total coefficient of determination (p 0.001) in a logisticregression study using 50% along with 25%/75% cut-off values. At the 50% level ofsignificance (2%vs. 7%; p 0.001), the relationship between ASHRO score andmortality rate was statistically significant.

Abstract

One deep learning method is the Convolutional Neural Network (CNN).Natural language processing problems like text classification are simplified using thisapproach. In this study, we use a deep learning strategy, namely the CNN method todeal with the issue of text classification. CNNs, which require a large deal of time aswell as finances to train and use, have been greatly impeded by the rise of Big Data andthe increased complexity of tasks. To get around these problems, we introduce aMapReduce-based CNN that rethinks what a CNN has learned by breaking it down intoa series of smaller networks and training them in parallel. Subsets of incoming text areanalysed by many autonomous networks.