Who and about What Speaks in “Cheerful” and “Sad” Texts: In Search of Discrimination Features in Texts of Different Emotional Tonalities

Abstract

This article focuses on the peculiarities of lexical and syntactical combinability of the Russian verb говорить (“to speak”) in Russian Internet texts of different emotion classes. The article aims to substantiate and validate the use of the established specific characteristics of the combinability of the lexeme as discriminant features serving to automatically detect eight emotional tonalities in Internet texts in Russian. The authors refer to a collection of texts found in the Подслушано ( The Overhead ) public page in the vk.com social network. Using the eight classes classification of emotions proposed by Lövheim, the researchers correlate each of the texts in their selection whose total volume is over a million tokens with a particular emotion by referring to the corresponding hashtags and the emotion mapping of the texts carried out by 36 assessors, Russian native speakers of 19–45 years old. The statistical analysis including term-frequency-inverse document frequency measure (TF-IDF) and analysis of lexeme frequency in eight sub-corpora proves that the Russian verb говорить does not have the same relevance in all sub-corpora, i.e. in four of them, it demonstrates a high relative frequency and a significant statistical specificity, but in the remaining four others it does not. Referring to the tools of corpus linguistics, the authors prove that to automatically attribute texts to a certain emotion class, it is essential to take into account the following peculiarities of lexical and syntactic combinability of the verb говорить : a high percentage of subjective syntactic connections, the frequency of particular lexemes (e.g. врач for the classes СТРАХ / УЖАС), and the total frequency of the lexemes belonging to one particular lexico-semantic group functioning as subject of the verbs; the frequency of separate collocations (e.g. когда люди говорят for the Злость / Гнев class); the frequency of separate syntaxemes (e.g. “ с собой / себе lemma [ говорить ]” for the ГРУСТЬ / Тоска class); the frequency of competing syntaxemes in the specific lexemes and collocations in the position of its subject, the frequency of the syntaxemes “lemma [ говорить ], что ” и “lemma [ говорить ]: (direct speech)”, marking the author’s proneness to focus on the content of what is being said in the form of direct and reported speech. After having been applied as parameters to run the classifier, the discriminate features increased the accuracy of classification for some emotion classes of texts.