Pronouns as machine learning markers in genre and gender text attribution

This article focuses on the statistical and machine learning approaches in sloving the tasks of automatic genre and gender text attribution using the forms of personal pronounsya [I], ty [you], my [we] as language markers. The research material was texts of discourses with a specific feature: spontaneous informal speech. Computer-mediated communication was represented by texts of several genres of the social network VKontakte: (a) personal written dialogues between men and women from VKontakte. The full text size was 114 046 words. The number of dialogue participants was 38 people (19 men, 19 women) aged 18 to 20; (b) 287 walls of the social network VKontakte with 9 951 001 words. All the participants were students of Tomsk State University. Transcriptions of oral texts were extracted from the RuTu-BiC database of the Russian speech corpus of Turkic-Russian bilinguals. The number of respondents was 138 people. The texts consist of 617 846 words. During the investigation, the authors used such methods as correlation analysis, the method of generalized linear models (GLM), criteria for testing statistical hypotheses (Wilcoxon, Kruskal-Wallis test), and machine learning. The analysis was implemented in the programming language R 4.0.5 using the quanteda library. The analysis was carried out in two stages: (1) the diagnostic power of the pronouns in the tasks of classification by genre forms; (2) gender opposition within genre forms of the texts. The authors proved the dependence of using groups of pronouns on the genre form of the text. The machine learning methods showed effectiveness of models using formal metrics and confirmed a significant degree of similarity in the use of pronouns in the texts of the VKontakte wall and VKontakte dialogues. These two types of communication are opposed to the genre forms of oral public communication. The studied groups of pronouns are better used for the classification of the genre of the text than for gender attribution. Gender differentiation is confirmed only in the texts of the VKontakte wall genre. The result of the “full dataset” is a classification within two genres (VKontakte dialogues combined with VKontakte Walls) and oral public communication. There is an actual significant increase in the accuracy of the classifier, which indicates the similarity of these two genres and their opposition to oral public communication in the binary classification. The results of the analysis show problems of automatic text classification based on the gender of the text’s author. It is necessary to pay attention to the genre form of the text. Such differences can be explained by genre features. Linguistic means of expressing intentions can affect the frequency of personal pronouns in the texts. The authors declare no conflicts of interests.

Keywords

Authors

References