TOOLS FOR SEMANTIC SEARCH AND ANSWER GENERATION IN UKRAINIAN-LANGUAGE MUSEUM SYSTEMS

Abstract

This article examines modern semantic search models and a search-based response generation approach for building intelligent museum information systems focused on Ukrainian-language and bilingual textual materials. It is shown that traditional lexical search, based on keyword matching, does not always ensure adequate quality in cultural heritage tasks, where relevance is determined not only by formal term matching but also by the semantic proximity of the query and the document. In this context, text vector representation models take on particular significance, as they allow texts to be represented in vector space and enable the retrieval of contextually relevant information even in the absence of direct lexical matches. This paper analyzes the LaBSE, paraphrase-multilingual-MiniLM-L12-v2, multilingual-E5-large-instruct, and BGE-M3 models in terms of Ukrainian language support, multilingualism, search accuracy, computational resource requirements, and suitability for local deployment. The generation of search-based responses is examined separately as an approach in which a generative model constructs a response based on fragments retrieved by a search module from an external knowledge base. It is argued that this approach is particularly suitable for museum information systems, as it reduces the risk of hallucinations, improves the factual accuracy of responses, and enables a source verification mechanism. The advantages and limitations of open and closed generative models in the context of Ukrainian-language museum services are also systematized. Based on the analysis, practical recommendations are formulated regarding the selection of search models and the configuration of answer generation systems for digital cultural heritage.