The verbal component of video game text as a resource for retraining large language models

Abstract

The study was conducted in the field of digital linguistics, whose objectives include studying the parameters of natural language processing. The problems of large language models with understanding communicative context are related to fundamental limitations of their nature, among the absence of genuine emotional intelligence, the complexity of interpreting multimodal signals, and difficulties with contextual understanding of situations. One way to overcome these limitations is to use a retraining method on specialised data sets. The article examines the verbal component of polycode video game text as a most important component of multimodal discourse, which determines the perception of events, characters and the nature of the player’s interaction with the virtual game world. Particular attention is paid to the specific characteristics of the linguistic material in video games: the complexity of the lexical, syntactic and stylistic organisation of game dialogues, emotional intensity, functional informativeness, and contextual variability. It is shown that game dialogues possess built-in annotations of emotions and pragmatic functions, which makes them a valuable resource for natural language processing tasks, including sentiment analysis, speech act recognition, and training of empathetic language models. Using fragments from the verbal component of the video game Disco Elysium, we demonstrate the implementation of emotional and contextual speech parameters of characters in the English and Russian versions. The importance of video game texts as a source of linguistic data for research in the field of corpus technologies and machine learning is noted. It is concluded that the verbal component of video games is not only an artistic tool for narrative organisation, but also a promising basis for the development of interactive interfaces and context-sensitive language models.