Clients’ Freely Written Assessment as the Source of Automatically Mined Opinions

Abstract

Measuring the quality of products or services, a challenging task is to reveal clients’ satisfaction or sentiment. As people have many opportunities to express their opinions using various on-line channels (e.g., discussions, microblogs, social networks), the question is whether such data might be used for this purpose. Information hidden in the data includes the reasons why people perceive products or services as good or bad, what are the reasons of clients’ satisfaction or dissatisfaction, or what affects their sentiment. However, having the needed large amounts of data, it is hardly possible to process it manually. This paper presents a method that aims at automatic discovery of sources of human feelings hidden in textual messages that clients produce. For a demonstration, messages having a form of freely written reviews containing subjective evaluation of medical services were used. During analysis of the data, clusters representing groups of the whole reviews (or individual sentences) with a certain requested degree of similarity were created in an unsupervised manner. Then, a decision tree classifier was trained in order to find attributes (words) of the reviews that were significant for assigning the reviews to the clusters. Because individual words were sometimes not informative enough they were subsequently used as a starting point for searching for frequent multi-word expressions. As a result, the list of multi-word phrases representing frequent and important sources of clients’ opinions was presented.