Automated Analysis of Evidentiality in Russian Media Discourse: Experience with Neural Network Models

This article addresses the challenge of analyzing evidentiality within media discourse by leveraging contemporary machine learning methods. The study’s relevance stems from the need to interpret implicit and context-dependent modal meanings that are inadequately captured using traditional approaches. The aim is to develop a methodology for automated analysis of evidentiality while accounting for its interaction with other modal categories. A corpus of English-language media texts serves as the empirical basis for this research. An original method based on word vector representation algorithms, combinatorial modeling, and recurrent neural networks is proposed. As a result, evidential operators have been identified and classified, along with their stable patterns of cohesion with markers representing authorization, approximation, perception, modality, negation, evaluation, personalization, persuasiveness, expressivity, emotionality, and temporality. Typical trajectories of polymodal unfolding of evidential statements are presented. Special attention is given to the linguo-synergetic approach, which considers modal meanings as elements of a multilevel self-organizing semantic system. This methodology demonstrates high efficiency of neural network techniques in tasks related to automated discourse analysis and linguistic expertise.

evidentiality, modal categories, neural network methods, BiLSTM, Word2Vec, language modeling