Seite - 41 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

paper implements the normalization method by Sklearn.preprocessing.MinMaxScaler class: xnew= x−xmin xmax−xmin (1) where xmin is theminimumvalueof the sampledata, and xmax is themaximumvalueof the sampledata. 4.4. FeatureandKeywordsExtraction 4.4.1. UnstructuredDataFeatureExtraction Theselectionof featureentriesand theirweights iscalled the featureextractionof target samples, and the advantages anddisadvantages of feature extractionwill directly affect theoperationeffectof themodel.Except thewordvectors,weextract twootherkindsof features:word frequencyandkeywords. Here the TF-IDF (TermFrequency-InverseDocument Frequency) [9] algorithm is used forword frequency analysis to evaluate the importance of a term in the news text corpus. The importance of aword increases proportionallywith the number of times it appears in a text, but at the same timedecreases inverselywith the frequency it appears in the corpus.The top30words are selectedas the input of recognitionmodels.Table5 lists the top4words. Table5. TF-IDFWordFrequencyAnalysis. No. TOPWords Weights 1 Year-on-year 0.075403657 2 Increase 0.064093650 3 Trillion 0.055898826 4 Interest rate 0.050473410 Sentiment analysis canclassify thepolarityof thenews text anddeterminewhether the expressed opinion is positive, negative or neutral. In this paper, we define four keywords groups, which are key words, pos words, neg words and nonwords. The keywords are the keywords extracted in a half-automatic way: first the keywords are selectedbymatching thekeynouns in thenews textwithfinancial dictionary; then they arecheckedandfilteredbystudents fromfinancialmajor; thepos wordsandnegwords are polar verbs and adjectives, with thewords of positively or negatively affecting the capitaladequacylevel.Thesewordsareextractedmanuallywithprofessionalknowledge in thefinancialfieldanda largeamountof readingonSinanews text; nonwordsare the privativewords, suchasnoandnone.Someof thekeywordsexamplesasshowninTable 6. If a news text contains keywords and positivewords, it was thought to be a positive news, andwas labeled as 1. Likewise, negative andneutral newswere labeled -1 and0 respectively.Newswouldbe thought tobeneutral if it containsnoneof thesewords. Y.Duetal. /Predicting the InterbankCapitalAdequacyLevelBasedonFinancialDataAnalysis 41