WIXLIB

View metadata, citation and similar papers at core.ac.ukbrought to you byCOREprovided by Loughborough University Institutional RepositorySentiment Analysis using KNIME: aSystematic Literature Review of Big DataLogisticsGary Graham1, Royston Meriton, Patrick HennellyAbstract - Text analytics and sentiment analysiscan help researchers to derive potentiallyvaluable thematic and narrative insights fromtext-based content such as industry reviews,leading OM and OR journal articles andgovernment reports. The classification systemdescribed here analyses the opinions of theperformance of various public and private,manufacturing, medical, service and retailorganizations in integrating big data into theirlogistics. It explains methods of data collectionand the sentiment analysis process forclassifying big data logistics literature usingKNIME. Finally, it then gives an overview of thedifferences and explores future possibilities insentiment analysis for investigating differentindustrial sectors and data sources.1. INTRODUCTIONBig data logistics can be defined as the modellingand analysis of (urban) transport and distributionsystems through large data sets created by GPS,cell phone and transactional data of companyoperations, combined with human generatedactivity (i.e. social media, public transport) [1]. Thedemands and requirements are literally changing ona daily basis with the innovation in technologieswith smart computing and big data. All types oforganization whose logistics operation functions ina big data environment will have to adapt tochanging customer demands. At the same time theywill need to exploit the availability of big datatechnology to improve their process andoperational capabilities.Big data requires firms to have more technicaland technological supports to handle the five V’s ofBig Data and analytics that is “Volume”, “Variety”,Veracity”, “Value” and “Velocity” [2]. However,with the growth of big data there is privacysurveillance and data misuse challenges [3].Organizations also face challenges around quality,comprehensiveness, collection and the analysis ofdata from various sources. Furthermore, big dataalso needs to be robust, accessible, andinterpretable if it is to provide organizations withmeaningful opportunities and solutions.The purpose of this paper therefore is to explorethe risks and challenges of the firm implementing“big data logistics” into their operations. Secondly,to investigate the opportunities that big dataprovides the firm to improve their logisticsperformance. This will be achieved through the textprocessing of 552 records containing industryreviews, leading OM and OR journal articles andgovernment reports. We will analyse the opinionsof the performance of various public and private,manufacturing, medical, service and retailorganizations in integrating big data (analytics) intotheir logistics.II. DATA MINING TOOLSData mining is the process of discoveringinteresting knowledge from large amounts of datastored in databases, data warehouses, or otherinformation repositories. Data mining has manyapplication fields such as marketing, business,science and engineering, economics, games andbioinformatics.Text mining or sentiment analysis [4] is theanalysis of data contained in a natural languagetext, which deals with the computation of opinion,sentiment and subjectivity in text. Sentimentanalysis refers to the use of natural languageprocessing, text analysis and computationallinguistics to identify and extract subjectiveinformation from the text documents. The basictask of sentiment analysis is to determine thepolarity of a given texts.Currently, many data mining and knowledgediscovery tools and software are available for everyone and different usage such as the WaikatoEnvironment for Knowledge Analysis (WEKA) [5]RapidMiner [4] and Orange [6]. These tools andsoftware1 provide a set of methods and algorithmsthat help in the better utilization of data andinformation available to users; including methodsand algorithms for data analysis, cluster analysis,genetic algorithms, nearest neighbour, datavisualization, regression analysis, decision trees,predictive analytics and, text mining.Text mining tools, are not without theirlimitations. They tend to use a relatively simpledictionary approach to identify associations. Thismeans they cannot identify novel or newly namedbig data science phrases and terms. Another1Corresponding author, Leeds University BusinessSchool, Maurice Keyworth Building, University ofLeeds, Leeds, LS2 9JT. T: 44 (0) 113 343 8557, E:g.graham@leeds.ac.uk.1Jovic et al [8] provide a detailed review through acomparative study of diverse collection of data miningtools.

limitation lies in its inability to extract context ormeaning from sentences or terms. Methods that useartificial intelligence (AI), word context or machinelearning (ML) methods could potentially improvethe current term identification system [10]. In spiteof these limitations we suggest methodologicallythat text mining is a useful starting point inbuilding a neutral coding system for theoreticalframing and concept development [11].The tasks of dictionary making and sentimentanalysis are done in this paper by the means ofKNIME [7], which we found to be a user-friendlygraphical workbench capable of entire analysisprocess. KNIME uses six different steps to processtexts: reading and parsing documents, named entityrecognition, filtering and manipulation, wordcounting and keyword extraction, transformationand visualization. The choice of KNIME camebecause of its straightforward graphical userinterface for data pre-processing, its intuitive dataflow user interface and from its powerful datamining elements. However, there are also severallimitations including its high memoryrequirements, the complexity of Lab nodes and itssyncretic analysis of text as there are limitedsemantic and pragmatic analysis functionsavailable.III. KNIME METHODThe KNIME text processing feature was designedand developed to read and process textual data, andtransform it into numerical data (document andterm vectors) in order to apply regular KNIME datamining nodes (for classification and clustering).This feature allows for the parsing of textsavailable in various formats (here we used .csv) asKNIME data cells stored in a data table. It is thenpossible to recognise and tag different kinds ofnamed entities such as with positive and negativesentiment, thus enrichening the documentssemantically.Furthermore, documents can be filtered (e.g. bythe stop word or named entity filters), stemmed bystemmers for various languages pre-processed inmany other ways. Frequencies of words can becomputed, keywords extracted and documents canbe visualised (e.g. tag clouds). To apply regularKNIME nodes to cluster or classify documentsaccording to their sentiment, they can betransformed into numerical vectors.searchable fields and several document types thatpermit the user to easily narrow their searching.Both databases sort the results by parameters suchas; first author, cites, relevance and etc. The RefineResults section in both databases allows the user toquickly limit or exclude results by author, source,year, subject area, document type, institutions,countries, funding agencies and languages. Theresulting documents provide a citation, abstract,and references at a minimum. Results may beprinted, e-mailed, or exported to a citationmanager. The results may also be reorganizedaccording to the needs of the researcher by simplyclicking on the headings of each column. Oursearch of “big data logistics” documents resulted in552 records being retrieved from a ten year periodfrom 2006 to 2016.The described data was then loaded intoKNIME with the File Reader node and processed.In this phase, only records in English languagewere collected. Language of the text is set toEnglish and all texts that have different languagevalues are filtered out, because English dictionaryapplied on reviews and posts written in otherlanguages would not give results.V. DICTIONARY BUILDINGDictionary built for sentiment analysis of thephrase “big data” as it is used with respect to theterm “logistics” was graded only as positive ornegative. Scoring or sentiment analysis of thephrase “big data logistics” is done on the positivenegative level, therefore the phrase was analysedon the word level, giving each word associatedwith it a positive or negative polarity. For instance,efficiency would be scored positive whilst riskswould be scored negatively.For this task, publicly available MPQAsubjectivity lexicon was used as a starting point forrecognizing contextual polarity [7], this wasexpanded with a big data vocabulary built from theauthors previous papers [3]. The existing dictionarycontaining of approximately 8000 words isexpanded to fit the needs for sentiment analysis in away that initial portion of sentences are collected,which are separated into single words with Bag ofWords processing. Unnecessary words such assymbols or web URLs are filtered out, and alluseful, big data specific words are graded andadded to the dictionary. For instance, “veracity”,“value”, “volume”, “variety” and “velocity”.IV. DATA COLLECTIONVI. DATA SCORINGWOS and Scopus are powerful databases whichprovide different searching and browsing options[9]. The search options in both databases are theStandard Basic and Advanced. There are differentThe records were analysed on the word level givinga positive or negative grade for a term connected toeach phrase. Whilst text analytics of documents isusually accomplished simply with phrases counters

and mean calculations, our analytics of isfrequency-driven. Two separate work flows weretherefore built, one for calculating frequency basedon a grade and category, and other one for positivenegative (sentiment) grading.Total score for each word is given bymultiplying TF*IDF value with attitude of a term.Attitude can have one of three values depending onthe word polarity; -1 for word with negativepolarity, 1 for word with positive polarity and 0for neutral words.Final weights, which now represent attitude ofeach document , are grouped on the level ofdocument and binned into three bins to give one ofthree final results for each term; positive, negativeor neutral (Fig. 3).Figure 2 Big data logistics sentimentsA. Big Data Record GradingTF*IDF (Term Frequency*Inverse DocumentFrequency) [7] method assigns non-binary weightsrelated on a number of occurrences of a word.Weighting exploits counts from a backgroundcorpus, which is a large collection of documents;the background corpus serves as indication of howoften a word may be expected to appear in anarbitrary text. TF*IDF calculation determines howrelevant a given word is in a particular document.Besides term frequency 𝑓𝑤, 𝑑 which equals thenumber of times word 𝑤 appears in a document,size of the corpus 𝐷 is also needed. Given adocument collection, a word 𝑤 and an individualdocument 𝑑 𝜖 𝐷, TF*IDF value can be calculated:T𝐹 𝐼𝐷𝐹𝑤 𝑓𝑤, 𝑑 log𝐷𝑓 𝑤𝑑(1)Figure 3 TF-IDF ProcessingVII RESULTSTag clouds were initially used to visualise ourinitial findings. A simple tag cloud presented inFigure 4 gives the most used words in the positive(left hand cloud) and negative used words (righthand cloud).