OPTIMIZATION OF THE PROCESS OF ASSESSMENT OF OPEN FEEDBACKS OF RESPONDENTS BY MEANS OF ARTIFICIAL INTELLIGENCE

In this work, the characteristics of the task of assessment of respondents' answers, the main types of answers to questions, ISSN 1997-9266. Вісник Вінницького політехнічного інституту. 2014. № 1 102 the advantages of using open-ended questions are emphasized, and the main problems of evaluating answers to such questions with a large number of respondents, taking into account their characteristics, are considered. The report discusses various methods of presenting a verbal response in vector form. Various metrics for calculating the proximity between individual answers are considered and a method for calculating the distance between respondents' answers based on the modified Jacquard distance is proposed. The proposed distance takes n-gram proximity into account. A mathematical model of the evaluation process is proposed in the form of an optimization problem, which consists in finding the minimum sequence of respondents' answers, which the expert needs to evaluate, so that the average distance between unrated answers and rated answers does not exceed the threshold value. A method of solving the optimization problem using artificial intelligence methods, namely clustering methods, has been developed. The best and worst case of input data for the problem are described. The main idea of the proposed method is the step-by-step clustering of unrated answers of respondents and the selection of a characteristic answer from the largest cluster for manual evaluation by an expert. The concept of the respondent's characteristic answer is formulated and the method of its selection based on the normal distribution law of distances within the cluster is developed. It is proposed to automatically evaluate unrated answers by finding the minimum distance to rated answers. The results of the conducted experimental studies, which consisted in the selection of the data set, the vectorization method, and the clustering method, are given. Comparative tables and graphs of the use of various methods are given. The conducted studies empirically proved the adequacy of the developed mathematical models and methods and their independence from the specifics of the context of questions and answers and the characteristics of the respondents, such as the country of origin or residence of the respondent, the language of communication, and personal characteristics. The developed models and methods can be used in distance learning systems as support for the teacher's work