Machine Learning Techniques for Multi-class Opinion Mining Tasks

<h1>Abstract</h1><p>This paper examines two approaches to the problem of opinion mining: an approach based on machine learning methods and an approach based on the use of sentiment-lexicons. The method of automated extraction words from the text that carry emotional value was described and implemented. Subsequent text classification method based on the obtained sentiment-lexicons was described and implemented.</p><p>For some standard machine learning techniques (such method is naive Bayes classification, support vector machines, Random Forest) it is proposed to use words from the obtained sentiment-lexicon as classification features.</p><p>Methods are tested on real data &ndash;feedback and comments of Internet resource imhonet.ru user&rsquo;s on three subject areas: books, movies and camera. The data were presented at the seminar on Russian Information Retrieval Evaluation Seminar (ROMIP).</p><p>The optimal size of a sentiment-lexicon (which is automatically constructed) is experimentally determined and is based on sentiment-value, according to which words are ranked in the dictionary. The experiments revealed the optimal value of the sentiment value, ​​over which the part of the dictionary is cut-off. This eliminates from consideration part of the vocabulary with low concentration of sentiment words.</p><p>The paper presents experiments aimed to compare the quality of constructed sentiment-lexicons for different domains and different sets of input data.</p><p>For machine learning methods indicators of quality were calculated in addition to average ones (such as Accuracy, Macro_Precision, Macro_Recall). That allows us to compare machine learning methods and method based on sentiment lexicons.</p><p>It is also possible to compare the machine-learning techniques, working with words from the dictionary as classification features, and the method of calculating the aggregated sentiment-based text value based on sentiment of the words in the text and belonging to the lexicon.&nbsp;Comparisons were made for both the average quality values ​​obtained testing machine learning techniques and sentiment approach, and for the separate methods.</p><p style="margin-left:35.4pt;">&nbsp;</p><p style="margin-left:35.4pt;"><strong>Key words: </strong>Opinion mining, machine learning, sentiment analysis, SVM, Random Forest, Naive Bayes, lexicon-based approach.&nbsp;</p>