Revealing Groups of Semantically Close Textual Documents by Clustering: Problems and Possibilities

Introduction

People and companies have many opportunities to express their opinions related to a wide variety of topics. The media used for such communication include personal web pages and blogs, social networks, discussion boards, e-mail, instant messages, and others. Various subjects can benefit from a high availability of information, which also demands bigger involvement, knowledge, information processing and decision making skills. Due to huge volumes of data that is often freely available for many different subjects there is a need for approaches that enable to use the data for decision making. Since most of the data is available in an unstructured textual form, disciplines focusing on this type of data have gained on their significance during the last few years (Miner at al., 2012).

Because of inadequate time and effort that would be needed in order to reveal the knowledge hidden in the data, the processing cannot be often done manually by humans. Instead, the application of computer based automated methods is a more desirable choice. This is enabled by the availability of increased computational speed and memory sizes of ordinary computers as well as by the development of new algorithms that are able to address various needs and problems. Instead of a traditional methodology employing human operators for reading the documents, statistical analysis, and data mining techniques based on the non-linguistic structure of the documents (Dini & Mazzini, 2010), intelligent computer-based analysis called text mining might arrive at new and unforeseen results.

Text mining is a branch of computer science that uses techniques from data mining, information retrieval, machine learning, statistics, natural language processing, and knowledge management (Berry & Kogan, 2010). The greatest potential of text mining applications is in the areas where large quantities of textual data are generated and collected. These areas include, besides others, categorization of newspaper articles or web pages, e-mail filtering, organization of a library, customer complaints (or feedback) handling, marketing focus group programs, competitive intelligence, market prediction, extraction of topic trends in text streams, discovering semantic relations between events, or customer satisfaction analysis (Cao et al., 2014; Koteswara Rao & Dey, 2011; Miner at al., 2012; Nassirtoussi, 2014; Weiss et al., 2010). Text mining involves tasks such as text categorization, term extraction, single- or multi-document document summarization, clustering, association rules mining, or sentiment analysis (Feldman & Sanger, 2007).

At the end of the last century, machine learning gained on its popularity and became a dominant approach to text mining (Sebastiani, 2002). Machine learning is a discipline that focuses on modification or adaptation of computer behavior based on the past experience (the data in this case) so the behavior gets better in the future. Such an adaptation depends on whether there is the right behavior specified. If there is, it means that there is a set of examples with correct answers (actions) provided. In this case we talk about supervised learning. During the learning process a computer tries to generalize the knowledge to be able to react correctly to all, even previously unseen inputs. When the correct responses are not provided, a computer tries to find some patterns based on similarities between the inputs. This approach is known as unsupervised learning (Marsland, 2009). The common goal of both approaches is to achieve accuracy comparable to that achieved by human experts.