Neural Text Classification Using Discourse Structure of the Text

This paper considers the text classification problem. We verify quality for two tasks: fact-checking task and categorization of quotes from an Internet forum corresponding to argumentation. Attention to the automatic verification of facts has significantly enhanced after the "Pizzagate" scandal in the US in 2016. A strong model for solving these and analogous problems still does not exist. The aforementioned tasks require the style-based approach to achieve the best results because of the restriction on using additional data. The proposed model effectively encodes parsed discourse trees due to the recursive neural net. The similar method had the finest performance for the sentiment analysis and semantic relatedness of two sentences task before. In the paper, the comparison with popular methods is given. Experiments illustrate that the suggested model is effective and reaches top results for the assigned tasks. In addition, it confirms that discourse analysis helps to obtain a much higher quality for the classification of longer texts in the case of Internet argumentation analysis.