Year: 2025
Type: article
Abstract: The article examines the features of the using neural networks as full-fledged scientific and reference sources. It is concluded that a significant obstacle to using neural networks in the above-mentioned aspect is their focus on text synthesis - filling out ready-made forms based on keywords without implementation of deep analysis with the identification of the hierarchy of the phenomena under study and the establishment of their more subtle features. In specific texts, this feature is manifested in the appearance of three types of hallucinations. Hallucinations of the first type (reminiscent) are based on associations with specific texts and facts. Hallucinations of the second type (free-associative) arise as a result of associations with words contained in the request. The third group (based on typical series) is characterized by filling positions in ready-made templates with nominations (often incorrectly chosen), which is clearly visible in the examples of guides created by neural networks and which became the main material for this article. Hallucinations of the first and second groups arise when processing widespread queries, while the third group is more often manifested when working with highly specialized, rare facts. The work emphasizes the importance of a critical approach to generated texts.
Source: Communication Studies
Author: Roman E. Telpov
Institution: Pushkin State Russian language Institute
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204885313