Stress Sentimental Analysis Using Machine learning (Reddit): A Review
Sharad Rajyaguru
Mahavidyalaya,
health related issues are one of the leading problems in the world. There are so many health issues that happen due to the extensive usage of social media. In this paper, we survey text mining with help of Reedit dataset. In this survey, paper Researchers applied different pre processing techniques as well as an applied different feature extraction techniques. Researchers applied preliminary supervised learning approaches for detecting stress, both neural and traditional, as well as an analysis of the data's complexity and diversity,
well as the features of each category.
1. Introduction
A country's well being is dependent on its citizens' health. Asuddenoutbreakofadiseasecansend peopleinto a panic. The importanceofearlydiseaseidentificationcannotbeoverstated.It'scrucialtokeepitfromspreading acrossaregion.Thehelp ofthepublic'suseofsocialmediaplatformsisincreasingrapidly.Userpostingscanbeofgreatassistancetothosewhowantto voicetheirideas.Keeptrackofdiseaseoutbreaksindifferentpartsoftheworld.Asevidencedbytheinclusionofmentalhealth in the Sustainable Development Goals; there has been a growing recognition of the critical role mental health plays in achievingglobaldevelopmentgoalsinrecentyears.Oneoftheprimarycausesofdisabilityisdepression.Suicideisthesecond highestcauseofdeathinpeopleaged15to29.Peoplewithseriousmentalillnessesdieconsiderablysoonerthantheyshould uptotwodecadesearlier asaresultofpreventablephysicalailments.[1]Evenmanymorepeoplesaresufferingfrommental stress according to the survey of TCOH (Delhi based org) show that more than 74% and 88% of Indians, respectively, experiencetensionandanxiety.[2]TheauthorcollectsdatasetsfromonemostpopularsocialmediaformsisReddit,Reasonfor selectingthisdatabecausethereisnotextlimitationinRedditdatasetlikeTwitter.Redditdatasetisopensourceandfreeto download. According to a poll report by Marsh India, the country's largest insurance broker, three out of five employees in India(59percent)reportedfeelingextremely,oderately,orslightlystressed daily,a higherlevel thanboththeworldwideand Asiaareanorms.
Fig.1. BlockDiagramStresssentimentanalysis.[4]
International Research Journal of
2. Dataset detailed
Redditisasocialmediawebsitewhereuser’spost post topic
communitiescalledsubreddits.Thisdatawascreatedby ElsbethTurcanandKathleenMcKeown.[3] Thesedatasetsaredividedintofivecategorieslikeabuse,PTSD,anxiety,Social,and Financial.TherearebothtypesofDatasetavailable.Traindataset,testingdataset.ThesizeoftheTrainis2675kbandthetest datasizeis668kb.Numofrawin2838&columnis116inTrainingdatasetandtestingdatasetnoraw716andcolumnisa 116.Therearealsomanycolumnsavailableinthisdataset.
>
>
REVIEW OF PRE PROCESSING TECHNIQUES
ElsbethTurcanandKathleenMcKeown [3]workedondataanalysisintwodifferentcategories.Likebydomain,bythelabel.
By domain:BydomainAuthorsappliedtoExamineeachdomain'svocabularypatternsusingonlyourtrainingdata,excluding unlabeled data, to expand our analysis to the label level. First, we utilize the Linguistic Inquiry and Word Count (LIWC), a lexicon basedtoolthatprovidesscoresforpsychologicallyrelevantcategoriessuchasmelancholyorcognitiveprocesses,asa proxyfortopicpredominanceandexpressionvariation.
Fig3: LexicalDiversitybyDomain[3]
By label: Authors conduct similar studies on data that has been classified as stressful or non stressful by the majority of annotators. We validate certain typical findings in the mental health literature, such as the usage of more first person pronouns in stressful data (possibly representing higher self focus) and more social words in non stressful data (maybe reflectingabettersocialsupportnetwork).MICHAELM.TADESSE,HONGFEILIN,BOXU,AND
LIANG YANG [4] worked before proceeding to the feature selection and training stage, we use NLP tools to pre process the dataset. To begin, we break the posts into Distinct
usingtokenization.Next,wedeleteallURLs, punctuations, and stopwords that,ifleftalone,couldresult in unpredictable results. Then we use stemming to reduce the words to their root formandgrouptermsthatarerelatedtogether.
Here,theAuthorappliedpre processingtechniqueremovingcommonwords,Wordslemmatizationtechniquewasapplied. [5]
Maryam Mohammed [7] Aldarwish, and Hafiz Farooq Ahmedcollecta datasetfromdifferent generals like Face book, live Journals,andTwitter.MaryamMohammedAl Darwish, HafizFarooq Ahmedhasapplied Rapid Miner wasused tocreate the PredictingDepressionModel.Themodel ismadeupofseveralprocessesthatareusedtotestclassifiers,SVMclassifiers,and the Naive Bayes classifier. The first operator is Select Attributes, which determines which attributes of the training dataset shouldberetainedandwhichshouldberemoved.ThesecondandthirdoperatorsaretheNominaltoText,whichchangesthe type ofselectednominalattributestotextandmapsallattribute values tocorresponding stringvalues; itis usedinboththe training and test datasets. The fourth and fifth processes are Process Documents, which are used in the training and test datasetstogeneratewordvectorsfromstringattributesandconsistoffouroperators.Tokenize, FilterStopwords,Transform Cases, and Stem are the four Process Document operators. Tokenize operators convert a document`s text into a sequence of tokens.
PriyankaArora.[8],ParulAroracollecteddatasetfromTwitter3754tweets,andtheybothappliedpre processingtechniques likethis
Case Folding: Allthewordsinthedatasetareconvertedtolowercase.Thefollowingdatapretreatmentprocedureshavebeen eliminatedduetoalackoffunctionality.Unnecessarypunctuations,Extrablankspaces,URLs,Hashtags
Replacing Emoticons: Emoticonsarereplacedwithtermstorepresentdifferentemotions,suchaspositiveemoticons(EPOS), negativeemoticons(ENEG),andneutralemoticons(ENEUT).
Removal of stopwords: Stop words are words that do not add meaningful content to the dataset (for example, pronouns, prepositions,andconjunctions).Asaresult,deletingthemreducesthesizeoftheobjectsinthetrainingandtestingsetgreatly.
Replacement of sentiment words: Words representing sentiments are replaced with keywords, such as POS for positive phrasesandNEGfornegativewords.
Mrs.B.IdaSeraphim,SubrotoDas,ApoorvRanjan [9]collectdata from Kaggleandother government portals,then afterthey workedon.
Data cleaning: removenullvaluesandunnecessaryrowsfromthedatasetafterwhichAuthorswilldo Min max normalization or standardization.
REVIEW OF FEATURE EXTRACTION TECHNIQUES
TheAuthors applied three feature extraction techniques is applied Lexical features, Syntactic features, social media features, andSyntacticfeatures. [3]
Lexical features: The full suite of 93 LIWC features; average, maximum, and lowest scores for pleasantness, activation, and imageryfromtheDictionaryofeffectinLanguage(DAL);andsentimentassessedusingthePatternsentimentcollection
Syntactic features: Part of speech unigrams and bigrams, the Flesch Kincaid Grade Level, and the Automated Readability Index.
Social media features: The post's UTC timestamp; the ratio of up votes to down votes on the post, where an up vote approximatelycorrelatestoa"like"reactionandadownvotetoa"dislike"reaction.Thepostsnetscore(karma)(determined byReddit,basedonnupvotesanddownvotes)herenminesnopost.
InthispaperAuthorsappliedthefeatureextractiontechniquesisn gram,liwc,LDA,[4]
Volume:
ISSN:2395
2395
N gram modeling. It's a characteristic for depression detection that'scommonlyutilized
N gram: Thefeatures from the postings are examined
theirmethodnameisunigram&bigrams. Example unigram
bigrams
mining.
Table: 1 Example n gram [4]
SL.No. Typeofn gram Generation grams
1 Unigram [“I”,”reside”,”in”,“And”]
2 Bigrams [“Ireside”,“inAnand”]
AndNLP,formodelingauthors,isusedto(if IDF)thefullformTermfrequency inversedocumentfrequency
Topic modeling: Whenthevalidationsetislimitedto70topics,theLDAmodel performsbest.Onlywordsthat appearinat least10postingsareconsideredforthetopicselection.Everypostistreatedasasingledocumentthatmustbetokenizedand stemmed.
LIWC: To conduct Authors study, weselected 68 out of 95 features based on psycholinguistic measurements and converted each depressive and non depressive post into numerical values. Authors can then calculate scores for three higher level categoriesbasedonstandardlanguagedimensions,psychologicalprocesses,andpersonalconcerns inthisfashion.Oneofthe mostimportantaspectsoflanguage is the normal linguistic processes. The LIWC
PsycholinguisticVocabularyPackageisacollectionofpsycholinguisticvocabulary.
Feature type Method NumberofselectedFeature N grams UnigramsBiggrams 3000 2736
LinguisticDimensions psychological processpersonalconcern process
LIC 68 Topmodeling LDA 70
Table: 2 Differenttypesofapproachestotextencodingmethods. [4]
Lizakonopelko [5] is applied as a feature extraction technique. "Term Frequency Inverse Document Frequency" is abbreviated as TF IDF. This is a method for calculating the number of words in a collection of documents. Authors usually assigneachwordascoretoindicateitsprominenceinthedocumentandcorpus.Inthefieldsofinformationretrievalandtext mining,thismethodiscommonlyemployed.
sohommajumder[6]isappliedinfeatureextractionasaBerttokenized Priyanka[8]Arora,ParulAroraapplieddifferentkindsof featureextractiontechniquesTokenization,Stemming,GramfeaturesSentimentExtractionsandPOSvector.
In this paper Mrs. B. Ida Seraphim, Subroto Das, Apoorv Ranjan [9] appliedexploratoryanalysisforfeature selection and the second technique applied Binary. Based on the suicide incidences per 100,000 people, Authors chose to undertake a binary classificationofthesuicidedata,assigninghigh/lowsuicideriskgroups.The'Risk'columnisaddedtothe"full"dataframeas anextracolumn.
Suicides<mean(Suicides) >lowrisk >class0 Suicides>mean(Suicides) >highrisk >class1
A review of different classifiers used
In the paper applied Elsbeth Turcan, Kathleen McKeown[3] Support Vector Machines (SVMs), logistic regression; Nave Bayes, Perceptron,anddecisiontreesareamongthenon neuralmodelswetryfirst.Theparametersaretweaked.Gridsearchanda 10 foldmultiplierwereusedforthesemodels.Cross validation,andacquireresultsforavarietyofsituationInputandfeature combinations. As well also applied CNN (Convolution Neural Network) & both a two layer bidirectional Gated Recurrent Neural Network (GRNN) We try training embeddings with random initialization and initializing using our domain specific Word2Vec embeddings, as well as concatenating the best feature set from our non neural experiments onto the Representationsaftertherecurrentandconvolution/poolinglayers.WhenapplyingBERTtoanytask,weapplyitstraighttoit, fine tuning the pre trained BERT base8 on our classification task for three epochs. The appendix contains the parameter settingsforourvariousmodels.
InthispaperMICHAELM.TADESSE,HONGFEILIN,BOXU,ANDLIANGYANG[4]appliedLR(logisticregression),Randomforest, Adabost,SVM,mlp.]Here lizakonopelko [5] appliedSVC(SupportVectorClassifier) algorithms.Sohommajumder[6] hasapplied machine learning algorithms like LR, KNN, SVM, Xg boosts, Ada boosts, RF, Decision tree classifiers, GBM(Gradient Boosting Machine),Cart [7]Inthispaper,theAuthorsappliedMachinelearningalgorithmsLikeSvm&Naïvebabieswiththeparameter ofRecall,Accuracy,andF1score.Inthispaperpriyankaarora,parurlarora [8] appliedSVM&MultinomialNaïveBayes.Inthis surveypaper.Inthispaper Author [9] applied Decision Tree, AdaBoost, XGBoost, and Neural Network is some of our Machine Learningtechniques.
Result&Discussion
Models Precision Recall F1score
Majoritybaseline 0.51611 1.0000 0.6808
CNN+features* 0.60230 0.8455 0.7035
CNN* 0.5840 0.9322 0.7182
GRNNw/attention+features* 0.6792 0.7859 0.7286
n grambaseline* 0.7249 0.7642 0.7441
n grams+features* 0.7474 0.7940 0.7700
LogRegw/pretrainedWord2Vec+features 0.7346 0.8103 0.7706
LogRegw/fine tunedBERTLM+features* 0.7704 0.8184 0.7937
LogRegw/domainWord2Vec+features* 0.7433 0.8320 0.7980
BERT base* 0.7518 0.8699 0.8065
Table 3:SupervisedResults[3]
Ourbestmodel achievesa 79.80F1 scoreonourtestset,comparabletothestate of the artpretrainedBERT basemodel.In this table, “features” always refers to our best performing feature set (≥ 0.4 absolute Pearson’s r). Models marked with a * showanExcellentimprovement
Overthemajoritybaseline(approximate randomization test,p<0.01).AuthorsapplieddifferentModels;thebestmodel isa logistic regressionclassifier withWord2Vec embeddings.Themodel achievess79.80F1scoresonourtestset. Comparedto themajoritybaseline,then grambaseline,andthepre trainedembeddingmodel,therewasaconsiderableimprovement.Our logistic regression classifier outperforms BERT base (approximate randomization test, p > 0.5) with the added benefits of betterinterpretabilityandlessintensivetraining.Furthermore,asexpected,domain specificwordembeddingstrainedonour unlabeled corpus (Word2Vec, BERT) outperform n grams or Our logistic regression classifier outperforms BERT base (approximate randomization test, p > 0.5) with the added benefits of better interpretability and less intensive training. Furthermore,asexpected,domain specificwordembeddingstrainedonourunlabeledcorpus(Word2Vec,BERT)outperform
n gramsorpre trainedembeddings,highlightingtheimportanceofdomainknowledgeinthistask. Embeddings,highlighting the importance of domain knowledge in this task. Although our basic deep learning models do not outperform the majority baselineaswellasourstandardsupervisedmodelsorBERT,theyconsistentlyexceedthemajoritybaseline.
InthisPaperAuthors [4] applieddifferentclassificationalgorithmswithdifferentfeatureextractiontechniques. Bigramsacquirethehighest accuracy,particularlywiththe SVM learning method, which achieves 80 percent accuracy and a 0.79F1score;followedbytheLIWCfeatureanRFmodel(78percent,0.84),andLDAwithLRtextclassifier(78percent,0.84). (77 percent, 0.83). Bigram works well when evaluating single features; however, it has severe drawbacks when considering combinedfeatures.Intermsofpredictivepower,LIWCoutperformsLDAasasinglefeature.Thehighestprecisionisachieved withtheRFmodel.SVM(74percent,0.74)andMLP(70percent,0.72)wereoutperformedbySVM(78percent,0.84)andMLP (70 percent, 0.72). In contrast to past research, LDA outperforms LIWC in this study. To supplement the N grams, we use LIWC and LDA. In our investigation, the LIWC+LDA+bigram and MLP neural network model had the best performance for detectingdepression.ItexcelsSVM(90percent,0.91),LR(89percent,0.89),andRF(85percent, 0.85) and AdaBoost (85 percent, 0.85) with its 91 percent accuracy and 0.93 F1 score (79 percent, 0.81). Furthermore, it outperformsanotherfeaturesetcontainingunigrams(LIWC,LDA,unigram)asthebesttechnique,withRFasthebeststrategy (83percent,0.84). The predictivecapacityfor greaterperformanceisconcealedin goodfeatureselectionsand manyfeature combinations,accordingtoourfindingsBigramsacquirethehighest
accuracy,particularlywiththeSVMlearningmethod,whichAchieves80percentaccuracyanda0.79F1score;followedbythe LIWC feature an RF model (78 percent, 0.84), and LDA with LR text classifier (78 percent, 0.84). (77 pcent, 0.83). Bigram workswellwhenevaluatingsinglefeatures;however,ithasseveredrawbackswhenconsideringcombinedfeatures.Interms ofpredictivepower,LIWCoutperformsLDAasasinglefeature.ThehighestprecisionisachievedwiththeRFmodel.SVM(74 percent, 0.74) and MLP (70 percent, 0.72) were outperformed by SVM (78 percent, 0.84) and MLP (70 percent, 0.72). In contrast to past research, LDA outperforms LIWC in this study. To supplement the N grams, we use LIWC and LDA. In our investigation, the LIWC+LDA+bigram and MLP neural network model had the best performance for detecting depression. It excelsSVM(90percent,0.91),LR(89percent, 0.89), RF (85 percent, 0.85) and AdaBoost (85 percent, 0.85) with its 91 percent accuracy and 0.93 F1 score (79 percent, 0.81). Furthermore, it outperforms another feature set containingunigrams(LIWC,LDA,unigram)asthebesttechnique,withRFasthebeststrategy(83percent,0.84).Thepredictive capacity for greater performance is
and many feature combinations, according to our
Logisticregression,XGBoost,
Afterthe analysis
the
algorithmsandNaïveBayes
Bayesgivesthelowestaccuracy. MaryamMohammedAldarwish,
Table:-5 performanceofsupervisedalgorithms
getsexcellentresultscomparedtoothers.
HereshowthetablePriyankaAroraandParulrora[8]appliedSVR,MultinomialNaïveBayesalgorithmappliedhereMultinomial NaïveBayesget 78% accuracyandSVRget 79%7.Performanceapplied Researchers[9] algorithms by Ad boost werethe most accurateofthefourmachinelearningmodels,withascoreof97.6%,followedbyxg boostwithascoreof96.35percent.The decisiontreehad91.67%accuracy,andthemapclassifierhad90%accuracy.
4. Conclusion
Weallareleaveningonthedigitalera.Weallarehabitedtoextensiveuseofsocialmedia;theresomanymentalstress related health issues arise. Now if we talk about this survey Different techniques, different datasets, different authors, different objectives,anddifferentresults.ThereisaPre processingstepisalwaysperformedtoenhancetheinputofsubsequentsteps. Researchonintermediatesteps.[3] HereResearchersidentifiedthattheirdeeplearningmodelsarenotworkinguptothemark in terms of performance. [4] Authors experiment reveals that the applied approaches perform quite well; yet, the absolute valuesofthemetricsindicatethatthisisa difficultundertakingthatmeritsadditionalinvestigation.Authorsbelievethatthis experiment will help to establish thefoundation for newmechanisms to beused invarious sectors ofhealthcare to evaluate depressionandrelatedcharacteristics. [7] AuthorsusedRapidMinertotesttwoclassifiers(SVMandNaveBayesClassifier)in a predating depression model. Using the same patients as before data that has been incorporated into the planned online application, as well as Using a training dataset, the data was manually categorized.2073 depressing posts and 2073 non depressedpostswereused.ThereForeachofthethree outcomes, performance hasbeen calculated.TheNaive Bayesresults, thesentimentresults,andtheSVMresult. [8] Thisresearchproposesanovelapproachtoidentifyinghealthtweetsforsadness and anxiety from all mixed tweets, allowing us to determine health status in the real world. It's a new platform for patient interaction, with many of them taking part in decision making for better healthcare treatment outcomes. For future work, a systemthatismoreefficientandaccuratethanthecurrentsystemcanbebuilt,andtheprojectcanbeexpandedtoaidothers. Toassesstheservicequalityofanyhealthcarefacilitythespreadofdiseasesandtheconsequencesofmedicinalproductsstudy theexperiencesofothers'tweetsThesystemiscapableofcouldbedevelopedtoincludemorethanonelanguage,forexample, Urdu,Punjabi,andHindiarethreelanguagesspokeninIndia. [9]AccordingtoResearchersGradientboosteddecisiontreesand neuralnetworks,forexample,routinelybeatotheralgorithmsandhadthehighestaccuracyandprecision.Chatbotshadapoor accuracy rate,butAIisstill inits infancy, andthereis more Work to be done in the field of human emotion recognition. My pointofviewRelated [3][4]isAuthorsappliedgoodfeatureextractiontechniquesandgetaccurateresultsinstressdetection& analysis.
References
[1] Awasthi, P. (2020, December 11). 74% of Indians suffering from stress; 88% reported anxiety amid Covid 19: Study.The Hindu BusinessLine. https://www.thehindubusinessline.com/news/variet y/74 of indians suffering from stress 88 reported anxiety amid covid 19 study/article33306490.ece
[2] Mentalhealth.(2019,December19).WHO.https://www.who.int/health topics/mental
International Research Journal of Engineering and Technology (IRJET)
2395
p ISSN:2395 0072
[3] Turcan, E. and McKeown, K., 2019. Dreaddit: A Reddit dataset for stress analysis in social media. arXiv preprint arXiv:1911.00133.
[4] Tadesse, M.M., Lin, H., Xu, B. and Yang, L., 2019. Detection of depression related posts in reddit social media forum. IEEEAccess,7,pp.44883 44893.
[5] L. (2021, July 2). mental health subreddits. Kaggle. https://www.kaggle.com/code/lizakonopelko/mental health subreddits
[6] S.(2021,July3).BERTtokenizerwith9models NLPstress analysis.Kaggle. https://www.kaggle.com/code/sohommajumder21/bert tokenizer with 9 models nlp stress analysis
[7] Aldarwish, M.M. and Ahmad, H.F., 2017, March. Predicting depression levels using social media posts. In 2017 IEEE 13thinternationalSymposiumonAutonomousdecentralizedsystem(ISADS)(pp.277 280).IEEE.
[8] Arora, Priyanka, and ParulArora. "Mining Twitter data for depression detection." 2019 International Conference on SignalProcessingandCommunication(ICSC).IEEE,2019.
[9] Mrs.B.IdaSeraphim,SubrotoDas,ApoorvRanjan,2021,AMachineLearningApproachtoAnalyzeand [10]Predict Suicide Attempts, INTERNATIONAL JOURNALOFENGINEERINGRESEARCH&TECHNOLOGY(IJERT) Volume 10,Issue04(April2021)