COMPARATIVE STUDY OF VERBAL HINT DETECTION: HUMANS AND LANGUAGE MODELS WITH DIFFERENT LEVELS OF TRAINING

Abstract

The article aims to evaluate effectiveness of Retrieval-Augmented Generation (RAG) and Chain of Thought (CoT) technologies in training large language models (LLMs) to detect verbal hints. The study integrates methods of discourse analysis, machine learning, and cognitive linguistics, along with experimental testing of models on a corpus of diplomatic texts. The results confirm that the implementation of RAG and CoT significantly enhances the ability of LLMs to identify implicit meaning structures in speech. However, the analysis also reveals some limitations of these methods, particularly in interpreting contextual hints. The study establishes the relationship between hints, indirect speech acts, and implicit communication, while identifying three key factors for their recognition: semantics, intention, and context. The findings contribute to further research in NLP, pragmatics, and artificial intelligence and can aid in developing more accurate and interpretable natural language processing models.