Method of analysis of emotional coloring of texts using graph convolutional neural networks

Abstract

In the work, graph convolutional neural networks were investigated as a promising direction for the analysis of the emotional coloring of texts. The constructions of emotions, their nature and essence are described in detail. The main methods of classification of emotions are indicated, such as binary classification of emotions, classification of emotions with several labels and multi-class classification of emotions. The principle of forming graph neural networks is outlined and it is emphasized that models built on graph neural networks usually do not take into account the semantic meaning of the text, which refers to the meaning determined by the relations between words in the sentence, which is important for the classification of the emotions of the analyzed text. The concept of semantic and syntactic analysis of the text is separated, the approaches to implementation are described. It is emphasized that the semantic representation reflects the content of the text in a fairly structured form with the selection of different representations: abstract meaning representation, universal conceptual cognitive annotation, bilexical semantic dependencies and universal decomposition semantics. It is noted that the parser model is a graph-based semantic parser that solves the problem of constituent parsing. The model levels are characterized and the principle of operation is given. Mathematically, an acyclic indicative graph of the system of emotional coloring of the text is presented, the feature matrix and the adjacency matrix are described. A model of a neural network that works with graph-structured data is presented, which consists of three main layers. An approach to training such a neural network is proposed. It is emphasized that the model uses the concatenation of embedded words and syntactic embedded elements as input data. As a result, a convolutional neural network was obtained for the analysis of emotional coloring of texts and its characteristics compared to other types of neural networks were analyzed. It was concluded that although the accuracy of the developed method is quite high, other architectures of neural networks are better suited for the task of sentiment analysis.