Year of Graduation
System of Multipurpose Analysis of Russian-language Text Messages Based on IBM InfoSphere Streams
School of Software Engineering
The System of Multipurpose Analysis of Russian-language Text Messages Based on IBM InfoSphere Streams was developed. The system is used for categorization and sentiment analysis of reviews on products and services of Sberbank of Russia. The developed system has passed all delivery and acceptance tests and fits client needs and expectations. Comparative analysis of available tools and methods for text mining was conducted. For the given task the most appropriate methods and techniques of text mining were chosen. The Naive Bayes classifiers were trained to predict sentiment of review, product (ATM, mobile banking, Sberbank Online) and product characteristics (availability, speed, service quality) mentioned in it.Besides InfoSphere Streams Java operators, the third-party components were integrated as native operators: АОТ - tool for natural text processing and Weka - tool for statistical analysis. The choice of the third-party tools was based on researches and comparative analysis. The following precision was achieved: 82%, 80%, 77% for sentiment analysis, product and characteristic prediction respectively. The method of representation features as syntax connected tokens was proposed and the results of classification was improved by 5-7% in average.Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank method proposed by Stanford University was adapted to russian language. The advantages and disadvantages of the method was reviewed. Keywords: text mining, text mining tools, deep learning, text clustering, text categorization.